{ "build_id": "lec06-a539b1c41cfd976b", "course": { "slug": "lecture06", "label": "Lecture 06", "title": "预训练、规模扩展与上下文学习", "date": "2026-10-26", "origin": "https://codingai-lec06.pages.dev", "audience": "public" }, "sequence": [ "OV01", "OV02", "OV03", "OV04", "OV05", "OV06", "OV07", "OV08", "OV09", "OV10", "AR01", "AR02", "AR03", "AR04", "AR05", "AR06", "AR07", "AR08", "AR09", "AR10", "VR01", "VR02", "VR03", "VR04", "VR05", "VR06", "RG02", "RG03", "RG04", "RG09", "RG10", "RG19", "RG23", "RG24", "RG25", "RG30", "RG32", "RC-H1", "RC-H2", "RC-H3", "RC-P1", "RC-P2", "RC-P3", "RC-P4", "RC-P5", "RC-P6", "RC-A1", "RC-A2", "RC-A3", "RC-A4", "RC-A5", "RC-A6", "RC-C1", "RC-C2", "RC-T1", "RC-T2", "RC-T3", "RC-V1", "RC-V2", "RC-F1", "RC-F2", "CL01", "CL02", "CL03", "CL04" ], "optional_pages": { "AM01": { "page": "AR01", "step": "b0" }, "AM02": { "page": "AR01", "step": "b0" }, "AM03": { "page": "AR02", "step": "b0" }, "AM04": { "page": "AR02", "step": "b3" }, "AM05": { "page": "AR03", "step": "b0" }, "AM06": { "page": "AR03", "step": "b2" }, "AM07": { "page": "AR04", "step": "b0" }, "AM08": { "page": "AR04", "step": "b2" }, "AM09": { "page": "AR05", "step": "b0" }, "AM10": { "page": "AR06", "step": "b0" }, "AM11": { "page": "AR06", "step": "b2" }, "AM12": { "page": "AR07", "step": "b0" }, "AM13": { "page": "AR07", "step": "b1" }, "AM14": { "page": "AR09", "step": "b0" }, "AM15": { "page": "AR08", "step": "b0" }, "AM16": { "page": "AR08", "step": "b0" }, "AM17": { "page": "AR10", "step": "b0" }, "AM18": { "page": "AR10", "step": "b0" }, "N11": { "page": "VR01", "step": "b0" }, "N12": { "page": "VR02", "step": "b0" }, "N13": { "page": "VR02", "step": "b1" }, "N14": { "page": "VR02", "step": "b2" }, "N15": { "page": "VR03", "step": "b0" }, "N16": { "page": "VR03", "step": "b1" }, "N17": { "page": "VR04", "step": "b0" }, "N18": { "page": "VR05", "step": "b0" }, "N19": { "page": "VR05", "step": "b1" }, "GF01": { "page": "VR06", "step": "b0" }, "GF02": { "page": "RG02", "step": "b0" }, "GF03": { "page": "RG03", "step": "b0" }, "GF04": { "page": "RG04", "step": "b0" }, "GF05": { "page": "RG04", "step": "b0" }, "GF06": { "page": "RG04", "step": "b0" }, "GF07": { "page": "RG04", "step": "b0" }, "GF08": { "page": "RG04", "step": "b0" }, "GF09": { "page": "RG09", "step": "b0" }, "GF10": { "page": "RG10", "step": "b0" }, "GF11": { "page": "RG04", "step": "b0" }, "GF12": { "page": "RG09", "step": "b0" }, "GF13": { "page": "RG10", "step": "b0" }, "GF14": { "page": "RG10", "step": "b0" }, "GF14A": { "page": "RG10", "step": "b0" }, "GF15": { "page": "RG04", "step": "b0" }, "GF16": { "page": "RG10", "step": "b0" }, "GF17": { "page": "RG09", "step": "b0" }, "GF18": { "page": "RG25", "step": "b0" }, "GF19": { "page": "RG19", "step": "b0" }, "GF20": { "page": "RG19", "step": "b0" }, "GF21": { "page": "RG19", "step": "b0" }, "GF22": { "page": "RG19", "step": "b0" }, "GF23": { "page": "RG23", "step": "b0" }, "GF24": { "page": "RG24", "step": "b0" }, "GF25": { "page": "RG25", "step": "b0" }, "GF26": { "page": "RG25", "step": "b0" }, "GF27": { "page": "RG19", "step": "b0" }, "GF28": { "page": "RG09", "step": "b0" }, "GF29": { "page": "RG04", "step": "b0" }, "GF30": { "page": "RG30", "step": "b0" }, "GF31": { "page": "RG30", "step": "b0" }, "GF32": { "page": "RG32", "step": "b0" }, "N20": { "page": "RC-H1", "step": "b0" }, "N21": { "page": "RC-H2", "step": "b0" }, "N22": { "page": "RC-H2", "step": "b0" }, "N23": { "page": "RC-H3", "step": "b0" }, "N24": { "page": "RC-P1", "step": "b0" }, "N25": { "page": "RC-P2", "step": "b0" }, "N26": { "page": "RC-P3", "step": "b0" }, "N27": { "page": "RC-P4", "step": "b0" }, "N28": { "page": "RC-P4", "step": "b0" }, "N29": { "page": "RC-P4", "step": "b0" }, "N30": { "page": "RC-P5", "step": "b0" }, "N31": { "page": "RC-P6", "step": "b0" }, "N32": { "page": "RC-P6", "step": "b1" }, "N33": { "page": "RC-A1", "step": "b0" }, "N34": { "page": "RC-A2", "step": "b0" }, "N35": { "page": "RC-A3", "step": "b0" }, "N36": { "page": "RC-A1", "step": "b0" }, "N37": { "page": "RC-A4", "step": "b0" }, "N38": { "page": "RC-A4", "step": "b2" }, "N39": { "page": "RC-A5", "step": "b1" }, "N40": { "page": "RC-A6", "step": "b0" }, "N41": { "page": "RC-C1", "step": "b0" }, "N42": { "page": "RC-C2", "step": "b0" }, "N43": { "page": "RC-C2", "step": "b1" }, "N44": { "page": "RC-T1", "step": "b0" }, "N45": { "page": "RC-T1", "step": "b0" }, "N46": { "page": "RC-T2", "step": "b0" }, "N47": { "page": "RC-T2", "step": "b1" }, "N48": { "page": "RC-T2", "step": "b3" }, "N49": { "page": "RC-T3", "step": "b0" }, "N50": { "page": "RC-T3", "step": "b0" }, "N51": { "page": "RC-T3", "step": "b1" }, "N52": { "page": "RC-V1", "step": "b0" }, "N53": { "page": "RC-V1", "step": "b0" }, "N54": { "page": "RC-V2", "step": "b0" }, "N55": { "page": "RC-V2", "step": "b0" }, "N56": { "page": "RC-V2", "step": "b0" }, "N57": { "page": "RC-V2", "step": "b0" }, "N58": { "page": "RC-V2", "step": "b0" }, "N59": { "page": "RC-V2", "step": "b2" }, "N60": { "page": "RC-F1", "step": "b0" }, "N61": { "page": "RC-F1", "step": "b0" }, "N62": { "page": "RC-F2", "step": "b0" } }, "legacy_routes": { "O01": { "states": { "b0": { "path": "pages/O01.html", "from": "N52", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/O01.html", "from": "N52", "returnStep": "b0", "hash": "b1" }, "b2": { "path": "pages/O01.html", "from": "N52", "returnStep": "b0", "hash": "b2" } } }, "O02": { "states": { "b0": { "page": "N37", "step": "b0" }, "b1": { "page": "N37", "step": "b1" }, "b2": { "page": "N37", "step": "b2" } } }, "P01": { "states": { "b0": { "page": "N24", "step": "b1" }, "b1": { "page": "N26", "step": "b0" }, "b2": { "page": "N26", "step": "b2" } } }, "P02": { "states": { "b0": { "page": "N30", "step": "b0" }, "b1": { "page": "N30", "step": "b2" }, "b2": { "page": "N32", "step": "b2" } } }, "P03": { "states": { "b0": { "path": "pages/P03.html", "from": "N30", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/P03.html", "from": "N30", "returnStep": "b2", "hash": "b1" }, "b2": { "path": "pages/P03.html", "from": "N30", "returnStep": "b2", "hash": "b2" } } }, "P04": { "states": { "b0": { "path": "pages/P04.html", "from": "N32", "returnStep": "b1", "hash": "b0" }, "b1": { "path": "pages/P04.html", "from": "N32", "returnStep": "b1", "hash": "b1" }, "b2": { "path": "pages/P04.html", "from": "N32", "returnStep": "b1", "hash": "b2" } } }, "P05": { "states": { "b0": { "path": "pages/P05.html", "from": "AM10", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/P05.html", "from": "AM10", "returnStep": "b2", "hash": "b1" }, "b2": { "path": "pages/P05.html", "from": "AM10", "returnStep": "b2", "hash": "b2" }, "b3": { "path": "pages/P05.html", "from": "AM10", "returnStep": "b2", "hash": "b3" }, "b4": { "path": "pages/P05.html", "from": "AM10", "returnStep": "b2", "hash": "b4" } } }, "P06": { "states": { "b0": { "path": "pages/P06.html", "from": "N37", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/P06.html", "from": "N37", "returnStep": "b0", "hash": "b1" } } }, "H01": { "states": { "b0": { "page": "N38", "step": "b0" }, "b1": { "page": "N39", "step": "b0" }, "b2": { "page": "N40", "step": "b3" } } }, "H02": { "states": { "b0": { "path": "pages/H02.html", "from": "N40", "returnStep": "b3", "hash": "b0" }, "b1": { "path": "pages/H02.html", "from": "N40", "returnStep": "b3", "hash": "b1" }, "b2": { "path": "pages/H02.html", "from": "N40", "returnStep": "b3", "hash": "b2" } } }, "H03": { "states": { "b0": { "path": "pages/H03.html", "from": "N60", "returnStep": "b1", "hash": "b0" }, "b1": { "path": "pages/H03.html", "from": "N60", "returnStep": "b1", "hash": "b1" }, "b2": { "path": "pages/H03.html", "from": "N60", "returnStep": "b1", "hash": "b2" }, "b3": { "path": "pages/H03.html", "from": "N60", "returnStep": "b1", "hash": "b3" } } }, "H04": { "states": { "b0": { "path": "pages/H04.html", "from": "N40", "returnStep": "b3", "hash": "b0" }, "b1": { "path": "pages/H04.html", "from": "N40", "returnStep": "b3", "hash": "b1" }, "b2": { "path": "pages/H04.html", "from": "N40", "returnStep": "b3", "hash": "b2" } } }, "H05": { "states": { "b0": { "path": "pages/H05.html", "from": "N60", "returnStep": "b1", "hash": "b0" }, "b1": { "path": "pages/H05.html", "from": "N60", "returnStep": "b1", "hash": "b1" }, "b2": { "path": "pages/H05.html", "from": "N60", "returnStep": "b1", "hash": "b2" }, "b3": { "path": "pages/H05.html", "from": "N60", "returnStep": "b1", "hash": "b3" } } }, "L01": { "states": { "b0": { "path": "pages/L01.html", "from": "N40", "returnStep": "b3", "hash": "b0" }, "b1": { "path": "pages/L01.html", "from": "N40", "returnStep": "b3", "hash": "b1" }, "b2": { "path": "pages/L01.html", "from": "N40", "returnStep": "b3", "hash": "b2" }, "b3": { "path": "pages/L01.html", "from": "N40", "returnStep": "b3", "hash": "b3" } } }, "L02": { "states": { "b0": { "path": "pages/L02.html", "from": "N40", "returnStep": "b3", "hash": "b0" }, "b1": { "path": "pages/L02.html", "from": "N40", "returnStep": "b3", "hash": "b1" }, "b2": { "path": "pages/L02.html", "from": "N40", "returnStep": "b3", "hash": "b2" } } }, "S01": { "states": { "b0": { "page": "N41", "step": "b0" }, "b1": { "page": "N41", "step": "b2" }, "b2": { "page": "N41", "step": "b4" } } }, "S02": { "states": { "b0": { "path": "pages/S02.html", "from": "N42", "returnStep": "b1", "hash": "b0" }, "b1": { "path": "pages/S02.html", "from": "N42", "returnStep": "b1", "hash": "b1" }, "b2": { "path": "pages/S02.html", "from": "N42", "returnStep": "b1", "hash": "b2" } } }, "S03": { "states": { "b0": { "page": "N42", "step": "b2" }, "b1": { "page": "N43", "step": "b0" }, "b2": { "page": "N43", "step": "b2" }, "b3": { "page": "N43", "step": "b3" } } }, "S04": { "states": { "b0": { "path": "pages/S04.html", "from": "N43", "returnStep": "b4", "hash": "b0" }, "b1": { "path": "pages/S04.html", "from": "N43", "returnStep": "b4", "hash": "b1" } } }, "S05": { "states": { "b0": { "path": "pages/S05.html", "from": "N43", "returnStep": "b4", "hash": "b0" }, "b1": { "path": "pages/S05.html", "from": "N43", "returnStep": "b4", "hash": "b1" } } }, "E01": { "states": { "b0": { "page": "N45", "step": "b0" }, "b1": { "page": "N44", "step": "b2" }, "b2": { "page": "N45", "step": "b2" } } }, "E02": { "states": { "b0": { "path": "pages/E02.html", "from": "N46", "returnStep": "b1", "hash": "b0" }, "b1": { "path": "pages/E02.html", "from": "N46", "returnStep": "b1", "hash": "b1" } } }, "E03": { "states": { "b0": { "path": "pages/E03.html", "from": "N47", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/E03.html", "from": "N47", "returnStep": "b0", "hash": "b1" }, "b2": { "path": "pages/E03.html", "from": "N47", "returnStep": "b0", "hash": "b2" } } }, "E04": { "states": { "b0": { "path": "pages/E04.html", "from": "N48", "returnStep": "b3", "hash": "b0" }, "b1": { "path": "pages/E04.html", "from": "N48", "returnStep": "b3", "hash": "b1" }, "b2": { "path": "pages/E04.html", "from": "N48", "returnStep": "b3", "hash": "b2" }, "b3": { "path": "pages/E04.html", "from": "N48", "returnStep": "b3", "hash": "b3" }, "b4": { "path": "pages/E04.html", "from": "N48", "returnStep": "b3", "hash": "b4" }, "b5": { "path": "pages/E04.html", "from": "N48", "returnStep": "b3", "hash": "b5" }, "b6": { "path": "pages/E04.html", "from": "N48", "returnStep": "b3", "hash": "b6" } } }, "E05": { "states": { "b0": { "path": "pages/E05.html", "from": "N46", "returnStep": "b2", "hash": "b0" } } }, "E06": { "states": { "b0": { "path": "pages/E06.html", "from": "N46", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/E06.html", "from": "N46", "returnStep": "b2", "hash": "b1" } } }, "E07": { "states": { "b0": { "path": "pages/E07.html", "from": "N49", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/E07.html", "from": "N49", "returnStep": "b2", "hash": "b1" }, "b2": { "path": "pages/E07.html", "from": "N49", "returnStep": "b2", "hash": "b2" } } }, "E08": { "states": { "b0": { "path": "pages/E08.html", "from": "N50", "returnStep": "b3", "hash": "b0" }, "b1": { "path": "pages/E08.html", "from": "N50", "returnStep": "b3", "hash": "b1" }, "b2": { "path": "pages/E08.html", "from": "N50", "returnStep": "b3", "hash": "b2" }, "b3": { "path": "pages/E08.html", "from": "N50", "returnStep": "b3", "hash": "b3" } } }, "E09": { "states": { "b0": { "path": "pages/E09.html", "from": "N51", "returnStep": "b3", "hash": "b0" }, "b1": { "path": "pages/E09.html", "from": "N51", "returnStep": "b3", "hash": "b1" } } }, "E10": { "states": { "b0": { "path": "pages/E10.html", "from": "N51", "returnStep": "b3", "hash": "b0" }, "b1": { "path": "pages/E10.html", "from": "N51", "returnStep": "b3", "hash": "b1" } } }, "I01": { "states": { "b0": { "page": "N33", "step": "b0" }, "b1": { "page": "N33", "step": "b1" }, "b2": { "page": "N33", "step": "b2" } } }, "I02": { "states": { "b0": { "path": "pages/I02.html", "from": "AM03", "returnStep": "b1", "hash": "b0" }, "b1": { "path": "pages/I02.html", "from": "AM03", "returnStep": "b1", "hash": "b1" }, "b2": { "path": "pages/I02.html", "from": "AM03", "returnStep": "b1", "hash": "b2" } } }, "I03": { "states": { "b0": { "path": "pages/I03.html", "from": "AM04", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/I03.html", "from": "AM04", "returnStep": "b0", "hash": "b1" } } }, "I04": { "states": { "b0": { "path": "pages/I04.html", "from": "AM13", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/I04.html", "from": "AM13", "returnStep": "b2", "hash": "b1" } } }, "I05": { "states": { "b0": { "page": "N34", "step": "b0" }, "b1": { "page": "N34", "step": "b1" }, "b2": { "page": "N34", "step": "b3" } } }, "V01": { "states": { "b0": { "page": "N52", "step": "b1" }, "b1": { "page": "N53", "step": "b2" } } }, "V02": { "states": { "b0": { "path": "pages/V02.html", "from": "N54", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/V02.html", "from": "N54", "returnStep": "b2", "hash": "b1" }, "b2": { "path": "pages/V02.html", "from": "N54", "returnStep": "b2", "hash": "b2" } } }, "V03": { "states": { "b0": { "page": "N57", "step": "b0" }, "b1": { "page": "N57", "step": "b2" }, "b2": { "page": "N58", "step": "b2" } } }, "V04": { "states": { "b0": { "page": "N55", "step": "b0" }, "b1": { "page": "N55", "step": "b1" }, "b2": { "page": "N55", "step": "b3" } } }, "V05": { "states": { "b0": { "path": "pages/V05.html", "from": "N56", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/V05.html", "from": "N56", "returnStep": "b2", "hash": "b1" }, "b2": { "path": "pages/V05.html", "from": "N56", "returnStep": "b2", "hash": "b2" } } }, "V06": { "states": { "b0": { "path": "pages/V06.html", "from": "N59", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/V06.html", "from": "N59", "returnStep": "b2", "hash": "b1" }, "b2": { "path": "pages/V06.html", "from": "N59", "returnStep": "b2", "hash": "b2" } } }, "V07": { "states": { "b0": { "path": "pages/V07.html", "from": "N60", "returnStep": "b1", "hash": "b0" } } }, "C01": { "states": { "b0": { "page": "N61", "step": "b1" }, "b1": { "page": "N61", "step": "b1" }, "b2": { "page": "N61", "step": "b2" }, "b3": { "page": "N61", "step": "b3" }, "b4": { "page": "N61", "step": "b3" } } }, "M01": { "states": { "b0": { "path": "pages/N01.html", "from": "AM18", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/N01.html", "from": "AM01", "returnStep": "b1", "hash": "b1" } } }, "M02": { "states": { "b0": { "path": "pages/N04.html", "from": "AM04", "returnStep": "b1", "hash": "b1" }, "b1": { "path": "pages/N04.html", "from": "AM04", "returnStep": "b0", "hash": "b2" }, "b2": { "path": "pages/N04.html", "from": "AM05", "returnStep": "b0", "hash": "b3" } } }, "M03": { "states": { "b0": { "path": "pages/N05.html", "from": "AM09", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/N05.html", "from": "AM09", "returnStep": "b2", "hash": "b2" }, "b2": { "path": "pages/N06.html", "from": "AM12", "returnStep": "b0", "hash": "b0" }, "b3": { "path": "pages/N06.html", "from": "AM13", "returnStep": "b1", "hash": "b2" } } }, "M04": { "states": { "b0": { "path": "pages/N07.html", "from": "AM10", "returnStep": "b1", "hash": "b2" }, "b1": { "path": "pages/N08.html", "from": "AM14", "returnStep": "b2", "hash": "b1" }, "b2": { "path": "pages/N10.html", "from": "AM17", "returnStep": "b2", "hash": "b0" } } }, "Q01": { "states": { "b0": { "page": "N20", "step": "b2" }, "b1": { "page": "N20", "step": "b3" }, "b2": { "page": "N21", "step": "b2" } } }, "Q02": { "states": { "b0": { "page": "N22", "step": "b0" }, "b1": { "page": "N22", "step": "b1" }, "b2": { "page": "N23", "step": "b2" } } }, "P07": { "states": { "b0": { "page": "N27", "step": "b0" }, "b1": { "page": "N27", "step": "b1" }, "b2": { "page": "N27", "step": "b2" }, "b3": { "page": "N28", "step": "b0" }, "b4": { "page": "N28", "step": "b2" } } }, "X01": { "states": { "b0": { "page": "N11", "step": "b0" }, "b1": { "page": "N11", "step": "b2" } } }, "X02": { "states": { "b0": { "page": "N13", "step": "b2" }, "b1": { "page": "N14", "step": "b1" }, "b2": { "page": "N14", "step": "b3" } } }, "X03": { "states": { "b0": { "page": "N15", "step": "b1" }, "b1": { "page": "N16", "step": "b0" }, "b2": { "page": "N16", "step": "b3" } } }, "X04": { "states": { "b0": { "page": "N18", "step": "b1" }, "b1": { "page": "N19", "step": "b0" }, "b2": { "page": "N19", "step": "b1" } } }, "G02": { "states": { "b0": { "page": "N46", "step": "b1" }, "b1": { "page": "N47", "step": "b0" }, "b2": { "page": "N47", "step": "b1" }, "b3": { "page": "N47", "step": "b2" } } }, "G03": { "states": { "b0": { "page": "N49", "step": "b1" }, "b1": { "page": "N51", "step": "b0" }, "b2": { "page": "N50", "step": "b3" }, "b3": { "page": "N51", "step": "b2" } } }, "N01": { "states": { "b0": { "path": "pages/N01.html", "from": "AM18", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/N01.html", "from": "AM01", "returnStep": "b1", "hash": "b1" } } }, "N02": { "states": { "b0": { "path": "pages/N02.html", "from": "AM03", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/N02.html", "from": "AM04", "returnStep": "b0", "hash": "b1" }, "b2": { "path": "pages/N02.html", "from": "AM02", "returnStep": "b0", "hash": "b2" } } }, "N03": { "states": { "b0": { "path": "pages/N03.html", "from": "AM04", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/N03.html", "from": "AM04", "returnStep": "b1", "hash": "b1" }, "b2": { "path": "pages/N03.html", "from": "AM03", "returnStep": "b1", "hash": "b2" } } }, "N04": { "states": { "b0": { "path": "pages/N04.html", "from": "AM03", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/N04.html", "from": "AM04", "returnStep": "b1", "hash": "b1" }, "b2": { "path": "pages/N04.html", "from": "AM04", "returnStep": "b0", "hash": "b2" }, "b3": { "path": "pages/N04.html", "from": "AM05", "returnStep": "b0", "hash": "b3" } } }, "N05": { "states": { "b0": { "path": "pages/N05.html", "from": "AM09", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/N05.html", "from": "AM09", "returnStep": "b1", "hash": "b1" }, "b2": { "path": "pages/N05.html", "from": "AM09", "returnStep": "b2", "hash": "b2" }, "b3": { "path": "pages/N05.html", "from": "AM10", "returnStep": "b0", "hash": "b3" } } }, "N06": { "states": { "b0": { "path": "pages/N06.html", "from": "AM12", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/N06.html", "from": "AM13", "returnStep": "b0", "hash": "b1" }, "b2": { "path": "pages/N06.html", "from": "AM13", "returnStep": "b1", "hash": "b2" }, "b3": { "path": "pages/N06.html", "from": "AM13", "returnStep": "b2", "hash": "b3" } } }, "N07": { "states": { "b0": { "path": "pages/N07.html", "from": "AM10", "returnStep": "b0", "hash": "b0" }, "b1": { "path": "pages/N07.html", "from": "AM10", "returnStep": "b0", "hash": "b1" }, "b2": { "path": "pages/N07.html", "from": "AM10", "returnStep": "b1", "hash": "b2" }, "b3": { "path": "pages/N07.html", "from": "AM10", "returnStep": "b2", "hash": "b3" } } }, "N08": { "states": { "b0": { "path": "pages/N08.html", "from": "AM14", "returnStep": "b1", "hash": "b0" }, "b1": { "path": "pages/N08.html", "from": "AM14", "returnStep": "b2", "hash": "b1" }, "b2": { "path": "pages/N08.html", "from": "AM18", "returnStep": "b0", "hash": "b2" }, "b3": { "path": "pages/N08.html", "from": "AM14", "returnStep": "b2", "hash": "b3" } } }, "N09": { "states": { "b0": { "path": "pages/N09.html", "from": "AM14", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/N09.html", "from": "AM14", "returnStep": "b2", "hash": "b1" }, "b2": { "path": "pages/N09.html", "from": "AM14", "returnStep": "b2", "hash": "b2" }, "b3": { "path": "pages/N09.html", "from": "AM17", "returnStep": "b1", "hash": "b3" } } }, "N10": { "states": { "b0": { "path": "pages/N10.html", "from": "AM17", "returnStep": "b2", "hash": "b0" }, "b1": { "path": "pages/N10.html", "from": "AM17", "returnStep": "b2", "hash": "b1" }, "b2": { "path": "pages/N10.html", "from": "AM18", "returnStep": "b1", "hash": "b2" } } }, "B01": { "path": "index.html", "hash": "course-path" }, "B02": { "path": "index.html", "hash": "course-path" } }, "pages": [ { "id": "O01", "title": "下载到的模型里，究竟有什么？", "topic": "pretraining", "states": 3, "source_ids": [ "A01" ], "main": false, "url": "classroom/O01.html", "reading": "pages/O01.html" }, { "id": "O02", "title": "从继续一段文字，到回应一个请求", "topic": "pretraining", "states": 3, "source_ids": [], "main": false, "url": "classroom/O02.html", "reading": "pages/O02.html" }, { "id": "P01", "title": "普通文本，怎样自己给出下一道题？", "topic": "pretraining", "states": 3, "source_ids": [ "A07" ], "main": false, "url": "classroom/P01.html", "reading": "pages/P01.html" }, { "id": "P02", "title": "同一个预测目标，能遇到哪些关系？", "topic": "pretraining", "states": 3, "source_ids": [ "A07", "P03" ], "main": false, "url": "classroom/P02.html", "reading": "pages/P02.html" }, { "id": "P03", "title": "有多少数据，不等于训练读了多少", "topic": "pretraining", "states": 3, "source_ids": [ "P03" ], "main": false, "url": "classroom/P03.html", "reading": "pages/P03.html" }, { "id": "P04", "title": "训练的收益，能带到另一项任务吗？", "topic": "pretraining", "states": 3, "source_ids": [ "P01" ], "main": false, "url": "classroom/P04.html", "reading": "pages/P04.html" }, { "id": "P05", "title": "为什么一个单元既响应猫，也响应车？", "topic": "pretraining", "states": 5, "source_ids": [ "B01" ], "main": false, "url": "classroom/P05.html", "reading": "pages/P05.html" }, { "id": "P06", "title": "预测得更好，怎样变成回答得更合适？", "topic": "pretraining", "states": 2, "source_ids": [ "A07", "A06" ], "main": false, "url": "classroom/P06.html", "reading": "pages/P06.html" }, { "id": "H01", "title": "仍然预测 token，SFT 改了什么？", "topic": "posttraining", "states": 3, "source_ids": [ "A06" ], "main": false, "url": "classroom/H01.html", "reading": "pages/H01.html" }, { "id": "H02", "title": "先比较回答，再学习怎样评分", "topic": "posttraining", "states": 3, "source_ids": [ "A06", "P04" ], "main": false, "url": "classroom/H02.html", "reading": "pages/H02.html" }, { "id": "H03", "title": "分数怎样改变下一次生成？", "topic": "posttraining", "states": 4, "source_ids": [ "A06", "P04" ], "main": false, "url": "classroom/H03.html", "reading": "pages/H03.html" }, { "id": "H04", "title": "已有偏好对，可以直接优化", "topic": "posttraining", "states": 3, "source_ids": [ "P05" ], "main": false, "url": "classroom/H04.html", "reading": "pages/H04.html" }, { "id": "H05", "title": "让答案接受规则核对", "topic": "posttraining", "states": 4, "source_ids": [ "P06" ], "main": false, "url": "classroom/H05.html", "reading": "pages/H05.html" }, { "id": "L01", "title": "一条小旁路，怎样改变大矩阵？", "topic": "adaptation", "states": 4, "source_ids": [ "A05", "A15", "P07" ], "main": false, "url": "classroom/L01.html", "reading": "pages/L01.html" }, { "id": "L02", "title": "存成4-bit，不等于全程4-bit计算", "topic": "adaptation", "states": 3, "source_ids": [ "A15", "P08" ], "main": false, "url": "classroom/L02.html", "reading": "pages/L02.html" }, { "id": "S01", "title": "同一笔计算，分给模型还是数据？", "topic": "scaling", "states": 3, "source_ids": [ "P09", "P10" ], "main": false, "url": "classroom/S01.html", "reading": "pages/S01.html" }, { "id": "S02", "title": "同样的训练，横轴决定在问什么", "topic": "scaling", "states": 3, "source_ids": [ "P09" ], "main": false, "url": "classroom/S02.html", "reading": "pages/S02.html" }, { "id": "S03", "title": "固定计算，看模型与数据怎样分配", "topic": "scaling", "states": 4, "source_ids": [ "P10" ], "main": false, "url": "classroom/S03.html", "reading": "pages/S03.html" }, { "id": "S04", "title": "70B，为什么值得与280B比较？", "topic": "scaling", "states": 2, "source_ids": [ "P10" ], "main": false, "url": "classroom/S04.html", "reading": "pages/S04.html" }, { "id": "S05", "title": "模型还要被调用很多次", "topic": "scaling", "states": 2, "source_ids": [ "P10" ], "main": false, "url": "classroom/S05.html", "reading": "pages/S05.html" }, { "id": "E01", "title": "模型文件能放下，训练现场为什么放不下？", "topic": "distributed-training", "states": 3, "source_ids": [ "A19", "P11" ], "main": false, "url": "classroom/E01.html", "reading": "pages/E01.html" }, { "id": "E02", "title": "四张卡，先分数据", "topic": "distributed-training", "states": 2, "source_ids": [ "A22", "A23" ], "main": false, "url": "classroom/E02.html", "reading": "pages/E02.html" }, { "id": "E03", "title": "把相同的状态分开保管", "topic": "distributed-training", "states": 3, "source_ids": [ "A19", "P11" ], "main": false, "url": "classroom/E03.html", "reading": "pages/E03.html" }, { "id": "E04", "title": "ZeRO-3：省下副本之后，怎样算完一层？", "topic": "distributed-training", "states": 7, "source_ids": [ "A19", "A22", "A23", "P11" ], "main": false, "url": "classroom/E04.html", "reading": "pages/E04.html" }, { "id": "E05", "title": "一个大矩阵，也能由多卡合算", "topic": "distributed-training", "states": 1, "source_ids": [ "A21", "A23" ], "main": false, "url": "classroom/E05.html", "reading": "pages/E05.html" }, { "id": "E06", "title": "不同层排成流水线", "topic": "distributed-training", "states": 2, "source_ids": [ "A21", "A23" ], "main": false, "url": "classroom/E06.html", "reading": "pages/E06.html" }, { "id": "E07", "title": "注意力的大表，未必需要完整写回", "topic": "flashattention-moe", "states": 3, "source_ids": [ "A16", "A24", "A26", "P12" ], "main": false, "url": "classroom/E07.html", "reading": "pages/E07.html" }, { "id": "E08", "title": "分块以后，分子与分母一起走", "topic": "flashattention-moe", "states": 4, "source_ids": [ "A16", "A24", "P12" ], "main": false, "url": "classroom/E08.html", "reading": "pages/E08.html" }, { "id": "E09", "title": "反向需要中间量，就按块重算", "topic": "flashattention-moe", "states": 2, "source_ids": [ "A16", "A24", "P12" ], "main": false, "url": "classroom/E09.html", "reading": "pages/E09.html" }, { "id": "E10", "title": "参数很多，每次只激活一部分", "topic": "flashattention-moe", "states": 2, "source_ids": [ "A20", "A08" ], "main": false, "url": "classroom/E10.html", "reading": "pages/E10.html" }, { "id": "I01", "title": "放入一个已完成示例，接下来怎样续写？", "topic": "context-memory", "states": 3, "source_ids": [ "P03", "A07", "A05" ], "main": false, "url": "classroom/I01.html", "reading": "pages/I01.html" }, { "id": "I02", "title": "先把“按内容读取”拆成一个小例子", "topic": "context-memory", "states": 3, "source_ids": [ "B02", "P13" ], "main": false, "url": "classroom/I02.html", "reading": "pages/I02.html" }, { "id": "I03", "title": "干扰来自其他也有相似度的内容", "topic": "context-memory", "states": 2, "source_ids": [ "B02", "A04" ], "main": false, "url": "classroom/I03.html", "reading": "pages/I03.html" }, { "id": "I04", "title": "让较强的匹配获得更多份量", "topic": "context-memory", "states": 2, "source_ids": [ "B02", "A29" ], "main": false, "url": "classroom/I04.html", "reading": "pages/I04.html" }, { "id": "I05", "title": "这次增加了输入，模型参数改了吗？", "topic": "context-memory", "states": 3, "source_ids": [ "B02", "P13", "A04" ], "main": false, "url": "classroom/I05.html", "reading": "pages/I05.html" }, { "id": "V01", "title": "读完输入，与逐步生成是两种节奏", "topic": "serving", "states": 2, "source_ids": [ "A14", "A10" ], "main": false, "url": "classroom/V01.html", "reading": "pages/V01.html" }, { "id": "V02", "title": "KV 不是比喻，是一张容量账本", "topic": "serving", "states": 3, "source_ids": [ "A13", "A14" ], "main": false, "url": "classroom/V02.html", "reading": "pages/V02.html" }, { "id": "V03", "title": "有人答完了，新请求为什么还在等？", "topic": "serving", "states": 3, "source_ids": [ "A10", "A13", "A14" ], "main": false, "url": "classroom/V03.html", "reading": "pages/V03.html" }, { "id": "V04", "title": "当前记忆，也要有地方放", "topic": "serving", "states": 3, "source_ids": [ "A14" ], "main": false, "url": "classroom/V04.html", "reading": "pages/V04.html" }, { "id": "V05", "title": "完全相同的前缀，只留一份", "topic": "serving", "states": 3, "source_ids": [ "A13", "A14", "W01", "W02" ], "main": false, "url": "classroom/V05.html", "reading": "pages/V05.html" }, { "id": "V06", "title": "长输入与持续出字，可以分开组织", "topic": "serving", "states": 3, "source_ids": [ "A08", "A10", "A11", "A12", "W03", "W04" ], "main": false, "url": "classroom/V06.html", "reading": "pages/V06.html" }, { "id": "V07", "title": "生成引擎，也影响训练成本", "topic": "serving", "states": 1, "source_ids": [ "P04", "P06", "A09" ], "main": false, "url": "classroom/V07.html", "reading": "pages/V07.html" }, { "id": "C01", "title": "回到同一个请求，能力怎样成为回答？", "topic": "context-memory", "states": 5, "source_ids": [ "P03", "P13", "P11" ], "main": false, "url": "classroom/C01.html", "reading": "pages/C01.html" }, { "id": "M01", "title": "同一套 Transformer，怎样承载更多经验？", "topic": "context-memory", "states": 2, "source_ids": [ "P13", "B02" ], "main": false, "url": "classroom/M01.html", "reading": "pages/M01.html" }, { "id": "M02", "title": "查询 A，为什么 B 和 C 也跟了进来？", "topic": "context-memory", "states": 3, "source_ids": [ "B02" ], "main": false, "url": "classroom/M02.html", "reading": "pages/M02.html" }, { "id": "M03", "title": "怎样让这次读取更有选择？", "topic": "context-memory", "states": 4, "source_ids": [ "B02" ], "main": false, "url": "classroom/M03.html", "reading": "pages/M03.html" }, { "id": "M04", "title": "训练留下的，与这次读到的，有什么不同？", "topic": "context-memory", "states": 3, "source_ids": [ "P13", "B02" ], "main": false, "url": "classroom/M04.html", "reading": "pages/M04.html" }, { "id": "Q01", "title": "已知的一整段输入，怎样一起计算？", "topic": "flashattention-moe", "states": 3, "source_ids": [ "A04" ], "main": false, "url": "classroom/Q01.html", "reading": "pages/Q01.html" }, { "id": "Q02", "title": "序列变长，多出来的成本在哪里？", "topic": "flashattention-moe", "states": 3, "source_ids": [ "P11" ], "main": false, "url": "classroom/Q02.html", "reading": "pages/Q02.html" }, { "id": "P07", "title": "只换 big 与 small，指代为什么就变了？", "topic": "pretraining", "states": 5, "source_ids": [], "main": false, "url": "classroom/P07.html", "reading": "pages/P07.html" }, { "id": "X01", "title": "图像也能成为输入序列", "topic": "vision-scaling", "states": 2, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/X01.html", "reading": "pages/X01.html" }, { "id": "X02", "title": "小数据上的赢家，扩大训练后还一样吗？", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_VIT", "VISION_CONVNEXT" ], "main": false, "url": "classroom/X02.html", "reading": "pages/X02.html" }, { "id": "X03", "title": "谁给这么多图像出训练题？", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_MAE" ], "main": false, "url": "classroom/X03.html", "reading": "pages/X03.html" }, { "id": "X04", "title": "更大的模型，需要能用好它的任务", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_MAE" ], "main": false, "url": "classroom/X04.html", "reading": "pages/X04.html" }, { "id": "G02", "title": "四张卡，怎样少保存重复的训练状态？", "topic": "distributed-training", "states": 4, "source_ids": [ "A19", "P11" ], "main": false, "url": "classroom/G02.html", "reading": "pages/G02.html" }, { "id": "G03", "title": "注意力的大表，可以少搬几次吗？", "topic": "flashattention-moe", "states": 4, "source_ids": [ "A16", "A24", "A26", "P12" ], "main": false, "url": "classroom/G03.html", "reading": "pages/G03.html" }, { "id": "N01", "title": "同一套 Transformer，怎样用上更多经验？", "topic": "context-memory", "states": 2, "source_ids": [], "main": false, "url": "classroom/N01.html", "reading": "pages/N01.html" }, { "id": "N02", "title": "先说清楚：用什么找，想拿回什么？", "topic": "context-memory", "states": 3, "source_ids": [], "main": false, "url": "classroom/N02.html", "reading": "pages/N02.html" }, { "id": "N03", "title": "一次匹配，怎样变成一份内容？", "topic": "context-memory", "states": 3, "source_ids": [], "main": false, "url": "classroom/N03.html", "reading": "pages/N03.html" }, { "id": "N04", "title": "装进三条以后，为什么读回了不止 A？", "topic": "context-memory", "states": 4, "source_ids": [], "main": false, "url": "classroom/N04.html", "reading": "pages/N04.html" }, { "id": "N05", "title": "把弱响应压下去，读出会怎样变？", "topic": "context-memory", "states": 4, "source_ids": [], "main": false, "url": "classroom/N05.html", "reading": "pages/N05.html" }, { "id": "N06", "title": "attention 怎样把同一组分数变成权重？", "topic": "context-memory", "states": 4, "source_ids": [], "main": false, "url": "classroom/N06.html", "reading": "pages/N06.html" }, { "id": "N07", "title": "把三张卡收回 MLP：哪些东西由训练留下？", "topic": "context-memory", "states": 4, "source_ids": [], "main": false, "url": "classroom/N07.html", "reading": "pages/N07.html" }, { "id": "N08", "title": "attention 读取的内容，是这次输入算出来的", "topic": "context-memory", "states": 4, "source_ids": [], "main": false, "url": "classroom/N08.html", "reading": "pages/N08.html" }, { "id": "N09", "title": "这个请求结束以后，什么会留下？", "topic": "context-memory", "states": 4, "source_ids": [], "main": false, "url": "classroom/N09.html", "reading": "pages/N09.html" }, { "id": "N10", "title": "结构有潜力，还要看到扩大训练的结果", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/N10.html", "reading": "pages/N10.html" }, { "id": "N11", "title": "把一张图变成一串可以计算的向量", "topic": "vision-scaling", "states": 4, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/N11.html", "reading": "pages/N11.html" }, { "id": "N12", "title": "这张图比较的，究竟是哪一次训练？", "topic": "vision-scaling", "states": 4, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/N12.html", "reading": "pages/N12.html" }, { "id": "N13", "title": "在这个条件下，做大 ViT 没有换来更好成绩", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/N13.html", "reading": "pages/N13.html" }, { "id": "N14", "title": "扩大训练以后，同一对模型的排序变了", "topic": "vision-scaling", "states": 4, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/N14.html", "reading": "pages/N14.html" }, { "id": "N15", "title": "只看这些碎片，模型要补回什么？", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/N15.html", "reading": "pages/N15.html" }, { "id": "N16", "title": "重建可以不完美，监督却可以很明确", "topic": "vision-scaling", "states": 4, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/N16.html", "reading": "pages/N16.html" }, { "id": "N17", "title": "训练时补图，之后拿走的是哪一部分？", "topic": "vision-scaling", "states": 4, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/N17.html", "reading": "pages/N17.html" }, { "id": "N18", "title": "先立参照：同一批训练图片，从头训练能做到什么？", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/N18.html", "reading": "pages/N18.html" }, { "id": "N19", "title": "先学重建，再做分类，大模型的收益更明显", "topic": "vision-scaling", "states": 4, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": false, "url": "classroom/N19.html", "reading": "pages/N19.html" }, { "id": "N20", "title": "同一组已知输入，可以一起做哪些运算？", "topic": "distributed-training", "states": 4, "source_ids": [], "main": false, "url": "classroom/N20.html", "reading": "pages/N20.html" }, { "id": "N21", "title": "整段在手，不等于每一行都能偷看答案", "topic": "distributed-training", "states": 3, "source_ids": [], "main": false, "url": "classroom/N21.html", "reading": "pages/N21.html" }, { "id": "N22", "title": "长度翻倍，哪一块工作会接近四倍？", "topic": "flashattention-moe", "states": 4, "source_ids": [], "main": false, "url": "classroom/N22.html", "reading": "pages/N22.html" }, { "id": "N23", "title": "算得快之外，还要装得下、搬得动", "topic": "flashattention-moe", "states": 4, "source_ids": [], "main": false, "url": "classroom/N23.html", "reading": "pages/N23.html" }, { "id": "N24", "title": "这段普通实验记录，怎样自己产生训练题？", "topic": "pretraining", "states": 3, "source_ids": [ "A07" ], "main": false, "url": "classroom/N24.html", "reading": "pages/N24.html" }, { "id": "N25", "title": "同一段文字，给每个位置一个后续目标", "topic": "pretraining", "states": 4, "source_ids": [ "A07" ], "main": false, "url": "classroom/N25.html", "reading": "pages/N25.html" }, { "id": "N26", "title": "把真实后词的概率，变成一次参数更新", "topic": "pretraining", "states": 3, "source_ids": [ "A07" ], "main": false, "url": "classroom/N26.html", "reading": "pages/N26.html" }, { "id": "N27", "title": "“太大”的到底是哪一个？", "topic": "pretraining", "states": 3, "source_ids": [ "A07" ], "main": false, "url": "classroom/N27.html", "reading": "pages/N27.html" }, { "id": "N28", "title": "只把 big 换成 small，解释为什么反过来？", "topic": "pretraining", "states": 3, "source_ids": [ "A07" ], "main": false, "url": "classroom/N28.html", "reading": "pages/N28.html" }, { "id": "N29", "title": "翻译时，这一句、这个前缀和过去训练各提供什么？", "topic": "pretraining", "states": 3, "source_ids": [ "A07" ], "main": false, "url": "classroom/N29.html", "reading": "pages/N29.html" }, { "id": "N30", "title": "目标相同，材料让模型反复遇见的关系不同", "topic": "pretraining", "states": 3, "source_ids": [ "P03" ], "main": false, "url": "classroom/N30.html", "reading": "pages/N30.html" }, { "id": "N31", "title": "先看任务：这条结论能从前提推出吗？", "topic": "pretraining", "states": 4, "source_ids": [ "A07" ], "main": false, "url": "classroom/N31.html", "reading": "pages/N31.html" }, { "id": "N32", "title": "预训练之后，能否把所得用于另一项任务？", "topic": "pretraining", "states": 3, "source_ids": [ "P01" ], "main": false, "url": "classroom/N32.html", "reading": "pages/N32.html" }, { "id": "N33", "title": "参数先不动：给一个完整任务示例", "topic": "context-memory", "states": 3, "source_ids": [ "P03" ], "main": false, "url": "classroom/N33.html", "reading": "pages/N33.html" }, { "id": "N34", "title": "固定的是 W，改变的是这次进入计算的 X", "topic": "context-memory", "states": 4, "source_ids": [ "P03" ], "main": false, "url": "classroom/N34.html", "reading": "pages/N34.html" }, { "id": "N35", "title": "示例的作用，要在完整翻译任务上看", "topic": "context-memory", "states": 3, "source_ids": [ "P03" ], "main": false, "url": "classroom/N35.html", "reading": "pages/N35.html" }, { "id": "N36", "title": "给同一份 AB 记录加一个示例，会教清什么？", "topic": "context-memory", "states": 3, "source_ids": [ "P03" ], "main": false, "url": "classroom/N36.html", "reading": "pages/N36.html" }, { "id": "N37", "title": "同样继续写 token，可以是在续文档，也可以是在回答", "topic": "posttraining", "states": 3, "source_ids": [ "A06" ], "main": false, "url": "classroom/N37.html", "reading": "pages/N37.html" }, { "id": "N38", "title": "把刚才的请求和回答，变成一条 SFT 样本", "topic": "posttraining", "states": 3, "source_ids": [ "A06" ], "main": false, "url": "classroom/N38.html", "reading": "pages/N38.html" }, { "id": "N39", "title": "输入不直接计分，为什么仍然必须计算？", "topic": "posttraining", "states": 4, "source_ids": [ "A06" ], "main": false, "url": "classroom/N39.html", "reading": "pages/N39.html" }, { "id": "N40", "title": "换一份材料，检验留下的是行为还是旧答案", "topic": "posttraining", "states": 4, "source_ids": [ "A06" ], "main": false, "url": "classroom/N40.html", "reading": "pages/N40.html" }, { "id": "N41", "title": "同一笔计算，怎样分给模型和材料？", "topic": "scaling", "states": 5, "source_ids": [ "P09", "P10" ], "main": false, "url": "classroom/N41.html", "reading": "pages/N41.html" }, { "id": "N42", "title": "模型翻倍，剩下的训练量减半", "topic": "scaling", "states": 3, "source_ids": [ "P10" ], "main": false, "url": "classroom/N42.html", "reading": "pages/N42.html" }, { "id": "N43", "title": "谷底随预算移动，不能只追求参数更多", "topic": "scaling", "states": 5, "source_ids": [ "P10" ], "main": false, "url": "classroom/N43.html", "reading": "pages/N43.html" }, { "id": "N44", "title": "一次更新，为什么不只需要模型权重？", "topic": "distributed-training", "states": 3, "source_ids": [ "A19", "P11" ], "main": false, "url": "classroom/N44.html", "reading": "pages/N44.html" }, { "id": "N45", "title": "16 GB 的模型，训练状态可以是 128 GB", "topic": "distributed-training", "states": 3, "source_ids": [ "A19", "P11" ], "main": false, "url": "classroom/N45.html", "reading": "pages/N45.html" }, { "id": "N46", "title": "四张卡分担数据，为什么仍各存一份？", "topic": "distributed-training", "states": 3, "source_ids": [ "A19", "P11" ], "main": false, "url": "classroom/N46.html", "reading": "pages/N46.html" }, { "id": "N47", "title": "长期分片保存，轮到一层再临时取齐", "topic": "distributed-training", "states": 3, "source_ids": [ "A19", "P11" ], "main": false, "url": "classroom/N47.html", "reading": "pages/N47.html" }, { "id": "N48", "title": "分开保管，也能共同更新同一个模型", "topic": "distributed-training", "states": 4, "source_ids": [ "A19", "P11" ], "main": false, "url": "classroom/N48.html", "reading": "pages/N48.html" }, { "id": "N49", "title": "同一张注意力大表，为什么要搬来搬去？", "topic": "flashattention-moe", "states": 3, "source_ids": [ "A16", "A24", "A26", "P12" ], "main": false, "url": "classroom/N49.html", "reading": "pages/N49.html" }, { "id": "N50", "title": "分块以后，分子和分母都要一起接着算", "topic": "flashattention-moe", "states": 4, "source_ids": [ "A16", "A24", "A26", "P12" ], "main": false, "url": "classroom/N50.html", "reading": "pages/N50.html" }, { "id": "N51", "title": "全部块走完，写回的是同一个注意力输出", "topic": "flashattention-moe", "states": 4, "source_ids": [ "A16", "A24", "A26", "P12" ], "main": false, "url": "classroom/N51.html", "reading": "pages/N51.html" }, { "id": "N52", "title": "训练结束后，这一次请求从哪里开始？", "topic": "serving", "states": 3, "source_ids": [ "A14", "A10" ], "main": false, "url": "classroom/N52.html", "reading": "pages/N52.html" }, { "id": "N53", "title": "新 token 读取旧 K/V，再把自己的 K/V 留下来", "topic": "serving", "states": 4, "source_ids": [ "A14", "A10" ], "main": false, "url": "classroom/N53.html", "reading": "pages/N53.html" }, { "id": "N54", "title": "一个请求的当前记忆，可以占到多少空间？", "topic": "serving", "states": 4, "source_ids": [ "A14", "A10" ], "main": false, "url": "classroom/N54.html", "reading": "pages/N54.html" }, { "id": "N55", "title": "逻辑上连续，物理上可以按块长出来", "topic": "serving", "states": 4, "source_ids": [ "A14" ], "main": false, "url": "classroom/N55.html", "reading": "pages/N55.html" }, { "id": "N56", "title": "相同前缀，可以接出两条不同的后续", "topic": "serving", "states": 4, "source_ids": [ "A13", "A14", "W01", "W02" ], "main": false, "url": "classroom/N56.html", "reading": "pages/N56.html" }, { "id": "N57", "title": "A 已经答完，C 为什么还在等待？", "topic": "serving", "states": 3, "source_ids": [ "A10", "A13", "A14" ], "main": false, "url": "classroom/N57.html", "reading": "pages/N57.html" }, { "id": "N58", "title": "每轮接纳新请求，把空出来的机会用上", "topic": "serving", "states": 4, "source_ids": [ "A10", "A13", "A14" ], "main": false, "url": "classroom/N58.html", "reading": "pages/N58.html" }, { "id": "N59", "title": "输入很长、输出很长、人很多，是三种不同压力", "topic": "serving", "states": 4, "source_ids": [ "A08", "A10", "A11", "A12", "W03", "W04" ], "main": false, "url": "classroom/N59.html", "reading": "pages/N59.html" }, { "id": "N60", "title": "生成候选，又会成为下一轮训练的材料", "topic": "serving", "states": 4, "source_ids": [ "P04", "P06", "A09" ], "main": false, "url": "classroom/N60.html", "reading": "pages/N60.html" }, { "id": "N61", "title": "这一次回答，哪些内容从哪里来？", "topic": "context-memory", "states": 4, "source_ids": [ "P03", "P04", "P11", "P12" ], "main": false, "url": "classroom/N61.html", "reading": "pages/N61.html" }, { "id": "N62", "title": "多了一项测量，结论应当怎样改变？", "topic": "context-memory", "states": 4, "source_ids": [ "P03", "P04" ], "main": false, "url": "classroom/N62.html", "reading": "pages/N62.html" }, { "id": "AM01", "title": "模型要用上更多知识，先要能承载、能读取", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM01.html", "reading": "pages/AM01.html" }, { "id": "AM02", "title": "写进去了，还要在需要时读得回来", "topic": "memory-capacity", "states": 5, "source_ids": [], "main": false, "url": "classroom/AM02.html", "reading": "pages/AM02.html" }, { "id": "AM03", "title": "先写入两组“线索—内容”", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM03.html", "reading": "pages/AM03.html" }, { "id": "AM04", "title": "读A时，为什么B也会出来？", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM04.html", "reading": "pages/AM04.html" }, { "id": "AM05", "title": "条目越多，干扰怎样累积？", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM05.html", "reading": "pages/AM05.html" }, { "id": "AM06", "title": "能改的地方有三个", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM06.html", "reading": "pages/AM06.html" }, { "id": "AM07", "title": "同样一个线性层，可以把两条都读对", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM07.html", "reading": "pages/AM07.html" }, { "id": "AM08", "title": "最优线性写入能延后干扰，仍然有边界", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM08.html", "reading": "pages/AM08.html" }, { "id": "AM09", "title": "弱匹配经过非线性，能少贡献一些", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM09.html", "reading": "pages/AM09.html" }, { "id": "AM10", "title": "比较线索、调节响应、汇总内容：认出MLP", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM10.html", "reading": "pages/AM10.html" }, { "id": "AM11", "title": "改变响应函数，容量曲线为什么会移动？", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM11.html", "reading": "pages/AM11.html" }, { "id": "AM12", "title": "指数怎样拉开强匹配和弱匹配？", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM12.html", "reading": "pages/AM12.html" }, { "id": "AM13", "title": "归一化以后，就是熟悉的softmax读取", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM13.html", "reading": "pages/AM13.html" }, { "id": "AM14", "title": "有的关联在参数里，有的条目来自这次输入", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM14.html", "reading": "pages/AM14.html" }, { "id": "AM15", "title": "注意力的检索容量，要连同实验条件一起读", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM15.html", "reading": "pages/AM15.html" }, { "id": "AM16", "title": "上下文变长，读取的锐度也值得重新考虑", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM16.html", "reading": "pages/AM16.html" }, { "id": "AM17", "title": "有承载机制，还要把知识学进去", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM17.html", "reading": "pages/AM17.html" }, { "id": "AM18", "title": "机制说得通以后，去视觉实验中看结果", "topic": "memory-capacity", "states": 3, "source_ids": [], "main": false, "url": "classroom/AM18.html", "reading": "pages/AM18.html" }, { "id": "GF01", "title": "分数提高了，内部究竟学出了什么？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF01.html", "reading": "pages/GF01.html" }, { "id": "GF02", "title": "猫和汽车，为什么会出现在一起？", "topic": "goodfire-internals", "states": 2, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF02.html", "reading": "pages/GF02.html" }, { "id": "GF03", "title": "全图类别不同，局部结构可能相通", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF03.html", "reading": "pages/GF03.html" }, { "id": "GF04", "title": "一个解释，要同时接住输入、读出与反方向", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF04.html", "reading": "pages/GF04.html" }, { "id": "GF05", "title": "汽车为什么会经过看起来像“狗”的单元？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF05.html", "reading": "pages/GF05.html" }, { "id": "GF06", "title": "模型更大，共享的姿态一定会被拆开吗？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF06.html", "reading": "pages/GF06.html" }, { "id": "GF07", "title": "同一个神经元，换显微镜会看到什么？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF07.html", "reading": "pages/GF07.html" }, { "id": "GF08", "title": "工具改进之外，网络本身也会影响观察", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF08.html", "reading": "pages/GF08.html" }, { "id": "GF09", "title": "混合不是只看终点：沿训练看它怎样长出来", "topic": "goodfire-internals", "states": 4, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF09.html", "reading": "pages/GF09.html" }, { "id": "GF10", "title": "这里争论的是“为什么共享”，不是“能表示多少”", "topic": "goodfire-internals", "states": 4, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF10.html", "reading": "pages/GF10.html" }, { "id": "GF11", "title": "神经元在放行什么，又挡住什么？", "topic": "goodfire-internals", "states": 4, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF11.html", "reading": "pages/GF11.html" }, { "id": "GF12", "title": "规模变了，区分的粒度也变了吗？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF12.html", "reading": "pages/GF12.html" }, { "id": "GF13", "title": "把一层从 480 个通道压到 10 个，会怎样？", "topic": "goodfire-internals", "states": 4, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF13.html", "reading": "pages/GF13.html" }, { "id": "GF14", "title": "没找回细曲线，究竟说明了什么？", "topic": "goodfire-internals", "states": 4, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF14.html", "reading": "pages/GF14.html" }, { "id": "GF14A", "title": "如果把瓶颈放到最后一层呢？", "topic": "goodfire-internals", "states": 2, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF14A.html", "reading": "pages/GF14A.html" }, { "id": "GF15", "title": "从石墙、屋顶，走到修道院响应", "topic": "goodfire-internals", "states": 5, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF15.html", "reading": "pages/GF15.html" }, { "id": "GF16", "title": "换成 SAE 字典，会多看见什么？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF16.html", "reading": "pages/GF16.html" }, { "id": "GF17", "title": "罕见、看错与故意教错，是同一回事吗？", "topic": "goodfire-internals", "states": 4, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF17.html", "reading": "pages/GF17.html" }, { "id": "GF18", "title": "删掉弱响应，丢的是哪种信息？", "topic": "goodfire-internals", "states": 4, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF18.html", "reading": "pages/GF18.html" }, { "id": "GF19", "title": "同样的追踪方法，能沿一句话找到答案的来路吗？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF19.html", "reading": "pages/GF19.html" }, { "id": "GF20", "title": "一个神经元的负侧，可能藏着真正的工作", "topic": "goodfire-internals", "states": 4, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF20.html", "reading": "pages/GF20.html" }, { "id": "GF21", "title": "名字很多，是否就必须打包进隐蔽空间？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF21.html", "reading": "pages/GF21.html" }, { "id": "GF22", "title": "模型更大以后：一张选例图能告诉我们多远？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF22.html", "reading": "pages/GF22.html" }, { "id": "GF23", "title": "先写下解释，再让没见过原例的一方找新输入", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF23.html", "reading": "pages/GF23.html" }, { "id": "GF24", "title": "看起来很像，却缺了关键条件：解释还成立吗？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF24.html", "reading": "pages/GF24.html" }, { "id": "GF25", "title": "“约九成可解释”，分母究竟是什么？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF25.html", "reading": "pages/GF25.html" }, { "id": "GF26", "title": "残差流：轨迹图怎样才有解释力？", "topic": "goodfire-internals", "states": 5, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF26.html", "reading": "pages/GF26.html" }, { "id": "GF27", "title": "Attention：看哪里，还要问搬了什么", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF27.html", "reading": "pages/GF27.html" }, { "id": "GF28", "title": "规模变大：同一耳朵家族怎样细化？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF28.html", "reading": "pages/GF28.html" }, { "id": "GF29", "title": "地图要落到同一个真实节点", "topic": "goodfire-internals", "states": 4, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF29.html", "reading": "pages/GF29.html" }, { "id": "GF30", "title": "把尚未解释的例子留在桌面上", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF30.html", "reading": "pages/GF30.html" }, { "id": "GF31", "title": "从可解释案例，走向可执行研究", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF31.html", "reading": "pages/GF31.html" }, { "id": "GF32", "title": "回到 MAE：表示怎样落成一轮计算？", "topic": "goodfire-internals", "states": 2, "source_ids": [ "GOODFIRE_FULL" ], "main": false, "url": "classroom/GF32.html", "reading": "pages/GF32.html" }, { "id": "OV01", "title": "Transformer 搭起来了，能力还需要怎样解释？", "topic": "pretraining", "states": 1, "source_ids": [], "main": true, "url": "classroom/OV01.html", "reading": "pages/OV01.html" }, { "id": "OV02", "title": "预训练、后续适配与使用，分别在改变什么？", "topic": "pretraining", "states": 1, "source_ids": [ "P03", "P15" ], "main": true, "url": "classroom/OV02.html", "reading": "pages/OV02.html" }, { "id": "OV03", "title": "预测下一个词，或恢复遮住的内容", "topic": "pretraining", "states": 1, "source_ids": [ "P15" ], "main": true, "url": "classroom/OV03.html", "reading": "pages/OV03.html" }, { "id": "OV04", "title": "Scaling Law 让我们看见投入与效果的关系", "topic": "scaling", "states": 1, "source_ids": [ "P09", "P10" ], "main": true, "url": "classroom/OV04.html", "reading": "pages/OV04.html" }, { "id": "OV05", "title": "给几个例子，模型这一次改变了什么？", "topic": "context-memory", "states": 1, "source_ids": [ "P03" ], "main": true, "url": "classroom/OV05.html", "reading": "pages/OV05.html" }, { "id": "OV06", "title": "“预测与压缩”的直觉，还留下什么问题？", "topic": "pretraining", "states": 1, "source_ids": [], "main": true, "url": "classroom/OV06.html", "reading": "pages/OV06.html" }, { "id": "OV07", "title": "第一问：写进去之后，怎样还能读回来？", "topic": "memory-capacity", "states": 1, "source_ids": [ "B02" ], "main": true, "url": "classroom/OV07.html", "reading": "pages/OV07.html" }, { "id": "OV08", "title": "第二问：它学到了什么，怎样才算有证据？", "topic": "goodfire-internals", "states": 1, "source_ids": [ "VISION_MAE", "GOODFIRE_FULL" ], "main": true, "url": "classroom/OV08.html", "reading": "pages/OV08.html" }, { "id": "OV09", "title": "第三问：算式成立，机器怎样把它算完？", "topic": "distributed-training", "states": 1, "source_ids": [], "main": true, "url": "classroom/OV09.html", "reading": "pages/OV09.html" }, { "id": "OV10", "title": "先看清承载与证据，再回到文本训练", "topic": "memory-capacity", "states": 1, "source_ids": [ "B02" ], "main": true, "url": "classroom/OV10.html", "reading": "pages/OV10.html" }, { "id": "AR01", "title": "写进去了，为什么还可能用不上？", "topic": "memory-capacity", "states": 3, "source_ids": [ "B02" ], "main": true, "url": "classroom/AR01.html", "reading": "pages/AR01.html" }, { "id": "AR02", "title": "读 A 时，为什么 B 也出来了？", "topic": "memory-capacity", "states": 5, "source_ids": [ "B02" ], "main": true, "url": "classroom/AR02.html", "reading": "pages/AR02.html" }, { "id": "AR03", "title": "可以扩大空间，也可以改写入和读取", "topic": "memory-capacity", "states": 3, "source_ids": [ "B02" ], "main": true, "url": "classroom/AR03.html", "reading": "pages/AR03.html" }, { "id": "AR04", "title": "同样大小的 W，可以把两条都读对", "topic": "memory-capacity", "states": 3, "source_ids": [ "B02" ], "main": true, "url": "classroom/AR04.html", "reading": "pages/AR04.html" }, { "id": "AR05", "title": "让弱匹配在汇总前少出力", "topic": "memory-capacity", "states": 3, "source_ids": [ "B02" ], "main": true, "url": "classroom/AR05.html", "reading": "pages/AR05.html" }, { "id": "AR06", "title": "从已经发生的运算中，认出 MLP", "topic": "memory-capacity", "states": 3, "source_ids": [ "B02", "P13" ], "main": true, "url": "classroom/AR06.html", "reading": "pages/AR06.html" }, { "id": "AR07", "title": "指数拉开比例，softmax给出分配", "topic": "memory-capacity", "states": 3, "source_ids": [ "B02" ], "main": true, "url": "classroom/AR07.html", "reading": "pages/AR07.html" }, { "id": "AR08", "title": "同一实验里，读取规则改变了多少？", "topic": "memory-capacity", "states": 2, "source_ids": [ "B02" ], "main": true, "url": "classroom/AR08.html", "reading": "pages/AR08.html" }, { "id": "AR09", "title": "参数没变，这次可读的内容可以变", "topic": "memory-capacity", "states": 3, "source_ids": [ "B02" ], "main": true, "url": "classroom/AR09.html", "reading": "pages/AR09.html" }, { "id": "AR10", "title": "把结构上的可能性，交给真实训练检验", "topic": "memory-capacity", "states": 2, "source_ids": [ "B02" ], "main": true, "url": "classroom/AR10.html", "reading": "pages/AR10.html" }, { "id": "VR01", "title": "同一张照片，改变进入模型的方式", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_VIT", "VISION_MAE" ], "main": true, "url": "classroom/VR01.html", "reading": "pages/VR01.html" }, { "id": "VR02", "title": "训练条件改变，同一对模型的排序也改变", "topic": "vision-scaling", "states": 4, "source_ids": [ "VISION_VIT" ], "main": true, "url": "classroom/VR02.html", "reading": "pages/VR02.html" }, { "id": "VR03", "title": "模型预测在这里，目标也在这里", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_MAE" ], "main": true, "url": "classroom/VR03.html", "reading": "pages/VR03.html" }, { "id": "VR04", "title": "误差更新参数，训练后的表示再转用", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_MAE" ], "main": true, "url": "classroom/VR04.html", "reading": "pages/VR04.html" }, { "id": "VR05", "title": "同模型的收益，要连同额外训练一起读", "topic": "vision-scaling", "states": 3, "source_ids": [ "VISION_MAE" ], "main": true, "url": "classroom/VR05.html", "reading": "pages/VR05.html" }, { "id": "VR06", "title": "分数之外，网络内部在区分什么？", "topic": "goodfire-internals", "states": 2, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/VR06.html", "reading": "pages/VR06.html" }, { "id": "RG02", "title": "猫和汽车，为什么会出现在一起？", "topic": "goodfire-internals", "states": 2, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG02.html", "reading": "pages/RG02.html" }, { "id": "RG03", "title": "全图类别不同，局部结构可能相通", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG03.html", "reading": "pages/RG03.html" }, { "id": "RG04", "title": "一个解释，能否接住两侧与两端？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG04.html", "reading": "pages/RG04.html" }, { "id": "RG09", "title": "混合不是只看终点：沿训练看它怎样长出来", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG09.html", "reading": "pages/RG09.html" }, { "id": "RG10", "title": "能承载多少，与为什么共享", "topic": "goodfire-internals", "states": 1, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG10.html", "reading": "pages/RG10.html" }, { "id": "RG19", "title": "从一段文字，追到一个下一词", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG19.html", "reading": "pages/RG19.html" }, { "id": "RG23", "title": "把解释交出去，再测一个新输入", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG23.html", "reading": "pages/RG23.html" }, { "id": "RG24", "title": "句子很像，关系还一样吗？", "topic": "goodfire-internals", "states": 3, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG24.html", "reading": "pages/RG24.html" }, { "id": "RG25", "title": "“约九成”，究竟解释了什么？", "topic": "goodfire-internals", "states": 1, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG25.html", "reading": "pages/RG25.html" }, { "id": "RG30", "title": "把尚未解释的例子留在桌面上", "topic": "goodfire-internals", "states": 2, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG30.html", "reading": "pages/RG30.html" }, { "id": "RG32", "title": "回到 MAE：表示怎样落成一轮计算？", "topic": "goodfire-internals", "states": 2, "source_ids": [ "VISION_MAE", "GOODFIRE_FULL" ], "main": true, "url": "classroom/RG32.html", "reading": "pages/RG32.html" }, { "id": "RC-H1", "title": "同一组表示，怎样变成可执行的运算？", "topic": "distributed-training", "states": 3, "source_ids": [], "main": true, "url": "classroom/RC-H1.html", "reading": "pages/RC-H1.html" }, { "id": "RC-H2", "title": "一起计算，不等于工作量不再增长", "topic": "flashattention-moe", "states": 3, "source_ids": [], "main": true, "url": "classroom/RC-H2.html", "reading": "pages/RC-H2.html" }, { "id": "RC-H3", "title": "算得完、装得下、搬得动", "topic": "flashattention-moe", "states": 3, "source_ids": [], "main": true, "url": "classroom/RC-H3.html", "reading": "pages/RC-H3.html" }, { "id": "RC-P1", "title": "普通记录，怎样自己产生训练题？", "topic": "pretraining", "states": 3, "source_ids": [ "A07" ], "main": true, "url": "classroom/RC-P1.html", "reading": "pages/RC-P1.html" }, { "id": "RC-P2", "title": "一组参数，面对一排前缀与目标", "topic": "pretraining", "states": 4, "source_ids": [ "A07" ], "main": true, "url": "classroom/RC-P2.html", "reading": "pages/RC-P2.html" }, { "id": "RC-P3", "title": "把后词概率，变成一次参数更新", "topic": "pretraining", "states": 3, "source_ids": [ "A07" ], "main": true, "url": "classroom/RC-P3.html", "reading": "pages/RC-P3.html" }, { "id": "RC-P4", "title": "只改一个词，关系判断为什么反过来？", "topic": "pretraining", "states": 4, "source_ids": [ "A07" ], "main": true, "url": "classroom/RC-P4.html", "reading": "pages/RC-P4.html" }, { "id": "RC-P5", "title": "目标相同，材料让模型反复遇到什么？", "topic": "pretraining", "states": 3, "source_ids": [ "P03" ], "main": true, "url": "classroom/RC-P5.html", "reading": "pages/RC-P5.html" }, { "id": "RC-P6", "title": "普通文本训练，能帮助另一项任务吗？", "topic": "pretraining", "states": 4, "source_ids": [ "A07", "P01" ], "main": true, "url": "classroom/RC-P6.html", "reading": "pages/RC-P6.html" }, { "id": "RC-A1", "title": "给同一份 AB 请求加一个示例", "topic": "context-memory", "states": 3, "source_ids": [ "P03" ], "main": true, "url": "classroom/RC-A1.html", "reading": "pages/RC-A1.html" }, { "id": "RC-A2", "title": "固定 W，改变 X，当前状态就会改变", "topic": "context-memory", "states": 3, "source_ids": [ "P03" ], "main": true, "url": "classroom/RC-A2.html", "reading": "pages/RC-A2.html" }, { "id": "RC-A3", "title": "示例有效吗？看独立的完整翻译评价", "topic": "context-memory", "states": 3, "source_ids": [ "P03" ], "main": true, "url": "classroom/RC-A3.html", "reading": "pages/RC-A3.html" }, { "id": "RC-A4", "title": "把期望的回应方式写进训练样本", "topic": "posttraining", "states": 4, "source_ids": [ "A06" ], "main": true, "url": "classroom/RC-A4.html", "reading": "pages/RC-A4.html" }, { "id": "RC-A5", "title": "可读取谁，与哪些目标计分，是两件事", "topic": "posttraining", "states": 4, "source_ids": [ "A06" ], "main": true, "url": "classroom/RC-A5.html", "reading": "pages/RC-A5.html" }, { "id": "RC-A6", "title": "换事实，检验行为而非旧答案", "topic": "posttraining", "states": 3, "source_ids": [ "A06" ], "main": true, "url": "classroom/RC-A6.html", "reading": "pages/RC-A6.html" }, { "id": "RC-C1", "title": "固定预算，模型与训练量怎样取舍？", "topic": "scaling", "states": 3, "source_ids": [ "P09", "P10" ], "main": true, "url": "classroom/RC-C1.html", "reading": "pages/RC-C1.html" }, { "id": "RC-C2", "title": "沿一条真实预算线，看哪里用得更好", "topic": "scaling", "states": 5, "source_ids": [ "P10" ], "main": true, "url": "classroom/RC-C2.html", "reading": "pages/RC-C2.html" }, { "id": "RC-T1", "title": "16 GB 权重，为什么训练状态会更多？", "topic": "distributed-training", "states": 3, "source_ids": [ "A19", "P11" ], "main": true, "url": "classroom/RC-T1.html", "reading": "pages/RC-T1.html" }, { "id": "RC-T2", "title": "分担数据，与分担保存责任，是两个动作", "topic": "distributed-training", "states": 4, "source_ids": [ "A19", "P11" ], "main": true, "url": "classroom/RC-T2.html", "reading": "pages/RC-T2.html" }, { "id": "RC-T3", "title": "相同的读取目标，能否少保存、少搬运？", "topic": "flashattention-moe", "states": 4, "source_ids": [ "A16", "A24", "A26", "P12" ], "main": true, "url": "classroom/RC-T3.html", "reading": "pages/RC-T3.html" }, { "id": "RC-V1", "title": "旧参数接收新请求，生成当前状态", "topic": "serving", "states": 4, "source_ids": [ "A14", "A10" ], "main": true, "url": "classroom/RC-V1.html", "reading": "pages/RC-V1.html" }, { "id": "RC-V2", "title": "权重能共享，上下文状态仍要付成本", "topic": "serving", "states": 3, "source_ids": [ "A14", "A10", "A13", "W01", "W02", "A08", "A11", "A12", "W03", "W04" ], "main": true, "url": "classroom/RC-V2.html", "reading": "pages/RC-V2.html" }, { "id": "RC-F1", "title": "同一条回答的事实、本领与执行从哪里来？", "topic": "context-memory", "states": 3, "source_ids": [ "P03", "P04" ], "main": true, "url": "classroom/RC-F1.html", "reading": "pages/RC-F1.html" }, { "id": "RC-F2", "title": "新测量与新约束，真的改变选择", "topic": "context-memory", "states": 4, "source_ids": [ "P03", "P04" ], "main": true, "url": "classroom/RC-F2.html", "reading": "pages/RC-F2.html" }, { "id": "CL01", "title": "记忆容量：存入与读回，要放在一起判断", "topic": "memory-capacity", "states": 1, "source_ids": [ "B02" ], "main": true, "url": "classroom/CL01.html", "reading": "pages/CL01.html" }, { "id": "CL02", "title": "学到了什么：让解释接受新证据", "topic": "goodfire-internals", "states": 1, "source_ids": [ "GOODFIRE_FULL" ], "main": true, "url": "classroom/CL02.html", "reading": "pages/CL02.html" }, { "id": "CL03", "title": "可执行性：能力收益要与资源代价一起看", "topic": "serving", "states": 1, "source_ids": [], "main": true, "url": "classroom/CL03.html", "reading": "pages/CL03.html" }, { "id": "CL04", "title": "从一次回答，走向搜索、推理与验证", "topic": "posttraining", "states": 1, "source_ids": [], "main": true, "url": "classroom/CL04.html", "reading": "pages/CL04.html" } ], "labs": { "lecture": "06", "notebooks": [ { "key": "budget-adaptation", "title": "LAB 01 · 训练预算、适配与分块归一化", "ipynb": "labs/notebooks/budget-adaptation.ipynb", "html": "labs/html/budget-adaptation.html", "deepnote_id": "9272b71f695c48f4a86d387733872604", "deepnote_url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-R1-20261020-6df914c0-6647-47d4-b50a-ffcae662bb3a/notebook/LAB-01-9272b71f695c48f4a86d387733872604?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=9272b71f695c48f4a86d387733872604&utm_term=get_notebook", "sha256": "94654b1c2168247983aab8458658562d5fb07741673a502a318b6824c96456ff", "original_sha256": "b9a2917e369bcbf4f0df549f092365676f6b1f22692aa2280562241bc736ae8c", "deepnote_sha256": "b9a2917e369bcbf4f0df549f092365676f6b1f22692aa2280562241bc736ae8c", "local_markdown_revision": "2026-10-20", "changed_markdown_cells": [ 0, 3, 15 ], "cloud_sync": "unchanged original; current local prose not synced" }, { "key": "associative-memory", "title": "LAB 02 · 按内容读记忆：相似度、干扰与选择性", "ipynb": "labs/notebooks/associative-memory.ipynb", "html": "labs/html/associative-memory.html", "deepnote_id": "042b1d2a35d245878dbff53755515872", "deepnote_url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-R1-20261020-6df914c0-6647-47d4-b50a-ffcae662bb3a/notebook/LAB-02-042b1d2a35d245878dbff53755515872?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=042b1d2a35d245878dbff53755515872&utm_term=get_notebook", "sha256": "fb1aace115a88cddcaa1b52d6394f7d8251d7f94e7f23c9411ef552c7953e9d8", "original_sha256": "f9fa2f8010f3da17f3d6902524ecc6d67d977f7c2844bbc7559c0c0efa9caba1", "deepnote_sha256": "f9fa2f8010f3da17f3d6902524ecc6d67d977f7c2844bbc7559c0c0efa9caba1", "local_markdown_revision": "2026-10-20", "changed_markdown_cells": [ 0 ], "cloud_sync": "unchanged original; current local prose not synced" }, { "key": "kv-serving", "title": "LAB 03 · KV、分页、前缀与请求调度", "ipynb": "labs/notebooks/kv-serving.ipynb", "html": "labs/html/kv-serving.html", "deepnote_id": "ae30640da295448fb2f86cc57549f421", "deepnote_url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-R1-20261020-6df914c0-6647-47d4-b50a-ffcae662bb3a/notebook/LAB-03-KV-ae30640da295448fb2f86cc57549f421?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=ae30640da295448fb2f86cc57549f421&utm_term=get_notebook", "sha256": "f698f8ce0768308feb085cbdd6d631bf08fbc1f6163811feaf349e65036cfdd9", "original_sha256": "f698f8ce0768308feb085cbdd6d631bf08fbc1f6163811feaf349e65036cfdd9", "deepnote_sha256": "f698f8ce0768308feb085cbdd6d631bf08fbc1f6163811feaf349e65036cfdd9", "local_markdown_revision": null, "changed_markdown_cells": [], "cloud_sync": "unchanged original; current local prose not synced" }, { "key": "icl192", "title": "LAB 04 · 固定参数的 ICL 对照：192 请求离线链路", "ipynb": "labs/notebooks/icl192.ipynb", "html": "labs/html/icl192.html", "deepnote_id": "cf0999fef7b84a1b92cbbd939c94e560", "deepnote_url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-R1-20261020-6df914c0-6647-47d4-b50a-ffcae662bb3a/notebook/LAB-04-ICL-192-cf0999fef7b84a1b92cbbd939c94e560?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=cf0999fef7b84a1b92cbbd939c94e560&utm_term=get_notebook", "sha256": "c8f3deac9b078e1d104440d701936b329b1bc9c73d7ce2860f1d21520358e840", "original_sha256": "1ce601733988da02bc7018cc8d989b31183e7dc33bf11cae99a1657cfd0040fe", "deepnote_sha256": "1ce601733988da02bc7018cc8d989b31183e7dc33bf11cae99a1657cfd0040fe", "local_markdown_revision": "2026-10-20", "changed_markdown_cells": [ 0, 10 ], "cloud_sync": "unchanged original; current local prose not synced" } ], "execution": "preserved original successful CPU outputs; no new execution on 2026-10-20", "deepnote_project_id": "6df914c0-6647-47d4-b50a-ffcae662bb3a", "deepnote_navigation_id": "5403fe2c162a4783ace024a2413e9160", "deepnote_navigation_url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/6df914c0-6647-47d4-b50a-ffcae662bb3a/notebook/5403fe2c162a4783ace024a2413e9160?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=5403fe2c162a4783ace024a2413e9160&utm_term=get_notebook", "deepnote_native_readback": "exact name/type/content/order matched: 4 labs 84 blocks + navigation 7 blocks", "cloud_execution": "not run", "student_copy_run": "not verified", "native_visual_layout": "not browser-verified", "historical_candidate": "R1", "local_revision": { "date": "2026-10-20", "scope": "Markdown/navigation cleanup only; all code cells and stored outputs unchanged", "cloud_sync": "not performed", "current_archive": "Lecture06-Labs-20261020.zip", "historical_archive": "Lecture06-R1-Labs.zip" } }, "current_labs": { "schema_version": 1, "project": { "id": "2ff4b87b-2904-4b9c-b9e9-bdbd72962691", "name": "Fall 26 - Lecture 06", "url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-2ff4b87b-2904-4b9c-b9e9-bdbd72962691?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=2ff4b87b-2904-4b9c-b9e9-bdbd72962691&utm_term=get_project" }, "notebooks": [ { "key": "CLS00", "id": "a96e110cfb0d4b28b79d41a01b714f92", "title": "CLS 00 · 第六讲实验导航", "url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-2ff4b87b-2904-4b9c-b9e9-bdbd72962691/notebook/CLS-00-a96e110cfb0d4b28b79d41a01b714f92?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=a96e110cfb0d4b28b79d41a01b714f92&utm_term=get_notebook", "summary": "先按课堂问题找到本册和活动，再核对准备条件与证据类型。", "sections": [] }, { "key": "CLS01", "id": "e2c2fce2f81b40d59356bfe7a63361b0", "title": "CLS 01 · 训练目标与资源账本", "url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-2ff4b87b-2904-4b9c-b9e9-bdbd72962691/notebook/CLS-01-e2c2fce2f81b40d59356bfe7a63361b0?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=e2c2fce2f81b40d59356bfe7a63361b0&utm_term=get_notebook", "summary": "对齐本讲的 8B 配置，分别计算训练目标、参数状态与资源账本。", "sections": [] }, { "key": "LAB01", "id": "e1f6f8c1ebd543c5ade548e0af61edc1", "title": "LAB 01 · 记忆：写入、读回与容量", "url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-2ff4b87b-2904-4b9c-b9e9-bdbd72962691/notebook/LAB-01-e1f6f8c1ebd543c5ade548e0af61edc1?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=e1f6f8c1ebd543c5ade548e0af61edc1&utm_term=get_notebook", "summary": "固定查询与目标，比较写入、读回、干扰、选择性和容量。", "sections": [ { "key": "M1", "title": "M1 · 同一查询与目标，写得更多会怎样？", "block_id": "90410919d6644f819f533e400dca3dae" }, { "key": "M2", "title": "M2 · 改写入，还是改读取？", "block_id": "96c41419751341a8a0b667aaa5a447f5" }, { "key": "M3", "title": "M3 · 归一化、选择性与有限容量扫描", "block_id": "61fb56544b374691b3eef5ed9d93bcbb" }, { "key": "M4", "title": "M4 · 参数不变，内容可变；选择性也会放大错误", "block_id": "22f75058ab9d4d329b48a19e57a978b9" } ] }, { "key": "LAB02", "id": "e6eecd2d72554ef4adacc0b797db4b32", "title": "LAB 02 · 视觉：学到了什么，怎样验证", "url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-2ff4b87b-2904-4b9c-b9e9-bdbd72962691/notebook/LAB-02-e6eecd2d72554ef4adacc0b797db4b32?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=e6eecd2d72554ef4adacc0b797db4b32&utm_term=get_notebook", "summary": "区分训练、表示与独立评测；用配对干预和原研究图证检验解释。", "sections": [ { "key": "V1", "title": "V1 · 训练目标下降，权重真的变了吗？", "block_id": "53d432e203fc440ebba39d9c09c21895" }, { "key": "V2", "title": "V2 · 同一留出集：重建改善，读出也改善吗？", "block_id": "09c58e6186c64aa6a3392245acbca957" }, { "key": "V3", "title": "V3 · 哪些单元被用到了？先选候选，再做配对干预", "block_id": "6ea12ebcb1a94634ad76ca46cfec6442" }, { "key": "V4", "title": "V4 · 原研究证据台：预测、揭示，再审读边界", "block_id": "12bf3a75f0404c55bbedc705a3f30485" }, { "key": "V5", "title": "V5 · 同一个受训表示，怎样落成矩阵计算？", "block_id": "b1269a5efcb447548fffc67787021016" } ] }, { "key": "LAB03", "id": "9280cd5ad74c43cba2c65cd705ff8dd6", "title": "LAB 03 · 文本：目标、数据与参数更新", "url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-2ff4b87b-2904-4b9c-b9e9-bdbd72962691/notebook/LAB-03-9280cd5ad74c43cba2c65cd705ff8dd6?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=9280cd5ad74c43cba2c65cd705ff8dd6&utm_term=get_notebook", "summary": "从前缀与目标到真实轻量 RNN 更新，再比较材料和计分位置。", "sections": [ { "key": "T1", "title": "T1 · 完整材料怎样产生多个预测目标？", "block_id": "74c6517c9b184283a41a43808dbd1e8f" }, { "key": "T2", "title": "T2 · 一次 backward 和优化器步，究竟改变了什么？", "block_id": "7b37d910f87943aebb396f3416ebb3c6" }, { "key": "T3", "title": "T3 · 预算相同，只改变反复出现的材料", "block_id": "87e602bc25c841da9cd1fbf3d3e288f8" }, { "key": "T4", "title": "T4 · 只给回应计分，会改变可读范围吗？", "block_id": "a10fcbd9065c405baa1fe8530af6a535" } ] }, { "key": "LAB04", "id": "8d38f8a6b20944129b6587b282ab38d2", "title": "LAB 04 · 上下文：固定参数的对照", "url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-2ff4b87b-2904-4b9c-b9e9-bdbd72962691/notebook/LAB-04-8d38f8a6b20944129b6587b282ab38d2?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=8d38f8a6b20944129b6587b282ab38d2&utm_term=get_notebook", "summary": "核对固定模型的输入与已保存输出，重新评分并检查结论边界。", "sections": [ { "key": "I1", "title": "I1 · 哪些东西在这次运行中没有改变？", "block_id": "98722ed87eac4c9f8f8e47a170892bab" }, { "key": "I2", "title": "I2 · 对照能回答什么？先核输入，再看输出", "block_id": "e17ab6d3843841b7a00d3da644aea46a" }, { "key": "I3", "title": "I3 · 改一次分析条件，看看结论是否稳固", "block_id": "24c067a20fa748269347e230a31b65ae" }, { "key": "I4", "title": "I4 · 可选新推理：只改示例顺序，留下新证据", "block_id": "6b375e33eec343c48ff98426e1c3c54d" } ] }, { "key": "LAB05", "id": "6cbbfbeaa7b941c3af8063f4ca5969f8", "title": "LAB 05 · 同一次更新，多个数据分片", "url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-2ff4b87b-2904-4b9c-b9e9-bdbd72962691/notebook/LAB-05-6cbbfbeaa7b941c3af8063f4ca5969f8?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=6cbbfbeaa7b941c3af8063f4ca5969f8&utm_term=get_notebook", "summary": "将同一次更新拆到多个数据分片，比较局部梯度、汇总与参数更新。", "sections": [ { "key": "D1", "title": "D1 · 同初值、同八条样本，怎样合回一次更新？", "block_id": "018523536049452face3504fb4c32022" }, { "key": "D2", "title": "D2 · 每片同样一条序列，分母为什么仍不相同？", "block_id": "c21fe00d75534e5382ec6df4f68927f6" }, { "key": "D3", "title": "D3 · 候选一起进入分母时，还能只合并局部loss吗？", "block_id": "a9494ac7c3eb42238d88fcdd1495bbec" } ] }, { "key": "LAB06", "id": "56ea4de667c74a84a7be00689c8de2d9", "title": "LAB 06 · 资源：预算、搬运与执行选择", "url": "https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/Fall-26-Lecture-06-2ff4b87b-2904-4b9c-b9e9-bdbd72962691/notebook/LAB-06-56ea4de667c74a84a7be00689c8de2d9?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=56ea4de667c74a84a7be00689c8de2d9&utm_term=get_notebook", "summary": "从训练预算与内存状态走到注意力搬运、请求缓存和受约束选择。", "sections": [ { "key": "R1", "title": "R1 · 固定计算预算，改变模型与训练量", "block_id": "596d044a097841e989ee1ddc5e52b551" }, { "key": "R2", "title": "R2 · 16 GB 权重，为什么会有128 GB训练状态？", "block_id": "99a7a488a4a347d0bf8e5a36bb31c3ff" }, { "key": "R3", "title": "R3 · 同一 Q/K/V，分块怎样保持同一个读出？", "block_id": "c443b475b8e246b781dada9fba82493f" }, { "key": "R4", "title": "R4 · 共享权重，为什么八份上下文仍占八份空间？", "block_id": "f9b67b005ffd42c2bcf2edd475f01766" }, { "key": "R5", "title": "R5 · 先列可行集，再作一次真正会翻转的选择", "block_id": "fbaa3d39251c4f6eb017472ab8dff76b" } ] } ], "page_mappings": [ { "pages": [ "AR01" ], "states": [ "b0", "b1", "b2" ], "targets": [ { "notebook": "LAB01", "section": "M1" } ], "summary": "固定查询、目标和记忆前缀，比较 M=1、16、64、256 的读回。这里是受控关联记忆计算。" }, { "pages": [ "RG24" ], "states": [ "b0", "b1", "b2" ], "targets": [ { "notebook": "LAB02", "section": "V4" } ], "summary": "在 V4 选择 EVIDENCE_CASE=\"RG24\"，先写预测，再揭示原研究的近失配证据；本活动不重新测量模型激活。" }, { "pages": [ "RC-P1" ], "states": [ "b0", "b1", "b2" ], "targets": [ { "notebook": "LAB03", "section": "T1" } ], "summary": "从同一 A/B 记录生成多组前缀与目标，区分教学切分和真实 tokenizer。" }, { "pages": [ "RC-P2" ], "states": [ "b0", "b1", "b2", "b3" ], "targets": [ { "notebook": "LAB03", "section": "T1" } ], "summary": "检查 input、target、causal mask 及 BOS/EOS 对齐，说明各位置怎样使用同一组参数。" }, { "pages": [ "RC-P3" ], "states": [ "b0", "b1", "b2" ], "targets": [ { "notebook": "LAB03", "section": "T2" } ], "summary": "核对一次真实轻量 RNN 更新前后的参数、loss 和目标概率；不把教学计分预设为实测结果。" }, { "pages": [ "RC-P4" ], "states": [ "b0", "b1", "b2", "b3" ], "targets": [ { "notebook": "LAB03", "section": "T3" }, { "notebook": "LAB04", "section": "I2" } ], "summary": "比较成对输入与实际输出；把教师参考答案、轻量训练和独立诊断分别看待。" }, { "pages": [ "RC-P5" ], "states": [ "b0", "b1", "b2" ], "targets": [ { "notebook": "LAB03", "section": "T3" }, { "notebook": "LAB06", "section": "R1" } ], "summary": "区分材料库存、不同 token 和实际训练消费量，检查固定预算下材料提出的依赖。" }, { "pages": [ "RC-P6" ], "states": [ "b0", "b1", "b2", "b3" ], "targets": [ { "notebook": "LAB03", "section": "T3" } ], "summary": "分别阅读 GPT-1 原表和课堂轻量模型结果，说明它们各自支持怎样的迁移判断。" }, { "pages": [ "RC-A1" ], "states": [ "b0", "b1", "b2" ], "targets": [ { "notebook": "LAB04", "section": "I1" }, { "notebook": "LAB04", "section": "I2" } ], "summary": "核对固定 checkpoint、输入和已保存输出。24 条真实 pilot、独立 6 条诊断与 192 条合成 fixture 分开。" }, { "pages": [ "RC-A2" ], "states": [ "b0", "b1", "b2" ], "targets": [ { "notebook": "LAB01", "section": "M4" }, { "notebook": "LAB04", "section": "I1" } ], "summary": "比较参数不变时的内容变化；将关联记忆构造与固定真实模型的记录分开。" }, { "pages": [ "RC-A3" ], "states": [ "b0", "b1", "b2" ], "targets": [ { "notebook": "LAB04", "section": "I3" } ], "summary": "重算已有记录，检查分母和对照条件；D−C 不能单独代表示例的因果效应。" }, { "pages": [ "RC-A4", "RC-A5" ], "states": [ "b0", "b1", "b2", "b3" ], "targets": [ { "notebook": "LAB03", "section": "T4" } ], "summary": "分别指出可读取的输入和计分位置：只给回应计分，不等于看不到用户输入。" }, { "pages": [ "RC-A6" ], "states": [ "b0", "b1", "b2" ], "targets": [ { "notebook": "LAB03", "section": "T4" }, { "notebook": "LAB04", "section": "I2" }, { "notebook": "LAB04", "section": "I3" } ], "summary": "区分修改权重与修改上下文，分别核对已有记录和能够支持的结论。" }, { "pages": [ "CL03" ], "states": [ "b0" ], "targets": [ { "notebook": "LAB06", "section": "R5" } ], "summary": "使用 R5 账本说明当前为何选 A，以及改变哪个约束会改变选择。" } ], "topic_mappings": { "memory-capacity": [ [ "LAB01", "M1" ], [ "LAB01", "M4" ] ], "goodfire-internals": [ [ "LAB02", "V4" ] ], "pretraining": [ [ "LAB03", "T1" ], [ "LAB03", "T2" ], [ "LAB03", "T3" ], [ "LAB06", "R1" ] ], "context-memory": [ [ "LAB01", "M4" ], [ "LAB04", "I1" ], [ "LAB04", "I2" ], [ "LAB04", "I3" ] ], "posttraining": [ [ "LAB03", "T4" ] ], "serving": [ [ "LAB06", "R4" ], [ "LAB06", "R5" ] ] } }, "topics": [ "posttraining", "distributed-training", "adaptation", "memory-capacity", "context-memory", "pretraining", "serving", "scaling", "flashattention-moe", "vision-scaling", "goodfire-internals" ] }