svd-safety-l2_basis_remove40 / compression.json
Jeesup's picture
None arm at None budget
6ba6e10 verified
Raw History Blame Contribute Delete
15.9 kB
{
"source": "/workspace/svd_safety/assets/models/Llama-2-7b-chat-hf",
"upstream": "TUDa-HWAI/Basis_Sharing",
"step": "basis_sharing",
"ratio_removed": 0.4,
"group_size": 2,
"nsamples": 256,
"seqlen": 2048,
"seed": 42,
"share_part": [
"v",
"k",
"q",
"up",
"gate"
],
"private_part": [
"down",
"o"
],
"factored": "/workspace/svd_safety/results/basis_sharing/cells/l2_basis_remove40/truncated/factored.pt",
"param_fraction": 0.5999091746275906,
"ranks": {
"0.q_proj": 1638,
"0.k_proj": 1638,
"0.v_proj": 1638,
"0.o_proj": 1228,
"0.gate_proj": 2072,
"0.up_proj": 2072,
"0.down_proj": 1791,
"1.q_proj": 1638,
"1.k_proj": 1638,
"1.v_proj": 1638,
"1.o_proj": 1228,
"1.gate_proj": 2072,
"1.up_proj": 2072,
"1.down_proj": 1791,
"2.q_proj": 1638,
"2.k_proj": 1638,
"2.v_proj": 1638,
"2.o_proj": 1228,
"2.gate_proj": 2072,
"2.up_proj": 2072,
"2.down_proj": 1791,
"3.q_proj": 1638,
"3.k_proj": 1638,
"3.v_proj": 1638,
"3.o_proj": 1228,
"3.gate_proj": 2072,
"3.up_proj": 2072,
"3.down_proj": 1791,
"4.q_proj": 1638,
"4.k_proj": 1638,
"4.v_proj": 1638,
"4.o_proj": 1228,
"4.gate_proj": 2072,
"4.up_proj": 2072,
"4.down_proj": 1791,
"5.q_proj": 1638,
"5.k_proj": 1638,
"5.v_proj": 1638,
"5.o_proj": 1228,
"5.gate_proj": 2072,
"5.up_proj": 2072,
"5.down_proj": 1791,
"6.q_proj": 1638,
"6.k_proj": 1638,
"6.v_proj": 1638,
"6.o_proj": 1228,
"6.gate_proj": 2072,
"6.up_proj": 2072,
"6.down_proj": 1791,
"7.q_proj": 1638,
"7.k_proj": 1638,
"7.v_proj": 1638,
"7.o_proj": 1228,
"7.gate_proj": 2072,
"7.up_proj": 2072,
"7.down_proj": 1791,
"8.q_proj": 1638,
"8.k_proj": 1638,
"8.v_proj": 1638,
"8.o_proj": 1228,
"8.gate_proj": 2072,
"8.up_proj": 2072,
"8.down_proj": 1791,
"9.q_proj": 1638,
"9.k_proj": 1638,
"9.v_proj": 1638,
"9.o_proj": 1228,
"9.gate_proj": 2072,
"9.up_proj": 2072,
"9.down_proj": 1791,
"10.q_proj": 1638,
"10.k_proj": 1638,
"10.v_proj": 1638,
"10.o_proj": 1228,
"10.gate_proj": 2072,
"10.up_proj": 2072,
"10.down_proj": 1791,
"11.q_proj": 1638,
"11.k_proj": 1638,
"11.v_proj": 1638,
"11.o_proj": 1228,
"11.gate_proj": 2072,
"11.up_proj": 2072,
"11.down_proj": 1791,
"12.q_proj": 1638,
"12.k_proj": 1638,
"12.v_proj": 1638,
"12.o_proj": 1228,
"12.gate_proj": 2072,
"12.up_proj": 2072,
"12.down_proj": 1791,
"13.q_proj": 1638,
"13.k_proj": 1638,
"13.v_proj": 1638,
"13.o_proj": 1228,
"13.gate_proj": 2072,
"13.up_proj": 2072,
"13.down_proj": 1791,
"14.q_proj": 1638,
"14.k_proj": 1638,
"14.v_proj": 1638,
"14.o_proj": 1228,
"14.gate_proj": 2072,
"14.up_proj": 2072,
"14.down_proj": 1791,
"15.q_proj": 1638,
"15.k_proj": 1638,
"15.v_proj": 1638,
"15.o_proj": 1228,
"15.gate_proj": 2072,
"15.up_proj": 2072,
"15.down_proj": 1791,
"16.q_proj": 1638,
"16.k_proj": 1638,
"16.v_proj": 1638,
"16.o_proj": 1228,
"16.gate_proj": 2072,
"16.up_proj": 2072,
"16.down_proj": 1791,
"17.q_proj": 1638,
"17.k_proj": 1638,
"17.v_proj": 1638,
"17.o_proj": 1228,
"17.gate_proj": 2072,
"17.up_proj": 2072,
"17.down_proj": 1791,
"18.q_proj": 1638,
"18.k_proj": 1638,
"18.v_proj": 1638,
"18.o_proj": 1228,
"18.gate_proj": 2072,
"18.up_proj": 2072,
"18.down_proj": 1791,
"19.q_proj": 1638,
"19.k_proj": 1638,
"19.v_proj": 1638,
"19.o_proj": 1228,
"19.gate_proj": 2072,
"19.up_proj": 2072,
"19.down_proj": 1791,
"20.q_proj": 1638,
"20.k_proj": 1638,
"20.v_proj": 1638,
"20.o_proj": 1228,
"20.gate_proj": 2072,
"20.up_proj": 2072,
"20.down_proj": 1791,
"21.q_proj": 1638,
"21.k_proj": 1638,
"21.v_proj": 1638,
"21.o_proj": 1228,
"21.gate_proj": 2072,
"21.up_proj": 2072,
"21.down_proj": 1791,
"22.q_proj": 1638,
"22.k_proj": 1638,
"22.v_proj": 1638,
"22.o_proj": 1228,
"22.gate_proj": 2072,
"22.up_proj": 2072,
"22.down_proj": 1791,
"23.q_proj": 1638,
"23.k_proj": 1638,
"23.v_proj": 1638,
"23.o_proj": 1228,
"23.gate_proj": 2072,
"23.up_proj": 2072,
"23.down_proj": 1791,
"24.q_proj": 1638,
"24.k_proj": 1638,
"24.v_proj": 1638,
"24.o_proj": 1228,
"24.gate_proj": 2072,
"24.up_proj": 2072,
"24.down_proj": 1791,
"25.q_proj": 1638,
"25.k_proj": 1638,
"25.v_proj": 1638,
"25.o_proj": 1228,
"25.gate_proj": 2072,
"25.up_proj": 2072,
"25.down_proj": 1791,
"26.q_proj": 1638,
"26.k_proj": 1638,
"26.v_proj": 1638,
"26.o_proj": 1228,
"26.gate_proj": 2072,
"26.up_proj": 2072,
"26.down_proj": 1791,
"27.q_proj": 1638,
"27.k_proj": 1638,
"27.v_proj": 1638,
"27.o_proj": 1228,
"27.gate_proj": 2072,
"27.up_proj": 2072,
"27.down_proj": 1791,
"28.q_proj": 1638,
"28.k_proj": 1638,
"28.v_proj": 1638,
"28.o_proj": 1228,
"28.gate_proj": 2072,
"28.up_proj": 2072,
"28.down_proj": 1791,
"29.q_proj": 1638,
"29.k_proj": 1638,
"29.v_proj": 1638,
"29.o_proj": 1228,
"29.gate_proj": 2072,
"29.up_proj": 2072,
"29.down_proj": 1791,
"30.q_proj": 1638,
"30.k_proj": 1638,
"30.v_proj": 1638,
"30.o_proj": 1228,
"30.gate_proj": 2072,
"30.up_proj": 2072,
"30.down_proj": 1791,
"31.q_proj": 1638,
"31.k_proj": 1638,
"31.v_proj": 1638,
"31.o_proj": 1228,
"31.gate_proj": 2072,
"31.up_proj": 2072,
"31.down_proj": 1791
},
"stage": "basis_recovery_finetuned",
"recovery_finetune": {
"protocol": "basis_sharing_coefficient_lora",
"script": "gitclones/SVD-LLM/utils/LoRA.py",
"data_path": "alpaca-cleaned (/workspace/svd_safety/data/alpaca_cleaned.json)",
"epochs": 2,
"learning_rate": 0.0001,
"batch_size": 64,
"micro_batch_size": 4,
"cutoff_len": 256,
"val_set_size": 2000,
"max_train_samples": "all",
"precision": "bf16 (upstream hard-codes fp16; see run_with_torch_compat)",
"is_methods_own_stage2": false,
"note": "Basis Sharing publishes its own LoRA (lora.py: wikitext, batch 1, peft default q/v coefficient targets). This cell uses the project's SVD-LLM alpaca recipe so recovery data is constant across compressors, on all seven coefficient types and never on a basis.",
"lora_r": 8,
"lora_alpha": 16,
"lora_dropout": 0.05,
"lora_targets": [
"down_proj",
"gate_proj",
"k_proj",
"o_proj",
"q_proj",
"up_proj",
"v_proj"
],
"trained": "Coefficient modules only; shared and private bases frozen",
"merge": "C' = C + (alpha/r) B_lora A_lora, then W = C' @ basis",
"mean_rel_delta_coefficient": 0.02975503941498963
},
"param_fraction_is_upper_bound": false,
"rel_delta_coefficient": {
"0.down_proj": 0.01789172552525997,
"0.gate_proj": 0.07389885932207108,
"0.up_proj": 0.05307356268167496,
"0.k_proj": 0.04400179162621498,
"0.o_proj": 0.015720030292868614,
"0.q_proj": 0.042313627898693085,
"0.v_proj": 0.023663686588406563,
"1.down_proj": 0.009344649501144886,
"1.gate_proj": 0.05698956921696663,
"1.up_proj": 0.04757563769817352,
"1.k_proj": 0.02117219939827919,
"1.o_proj": 0.015290400013327599,
"1.q_proj": 0.02371150627732277,
"1.v_proj": 0.016866125166416168,
"2.down_proj": 0.017618203535676003,
"2.gate_proj": 0.05127978324890137,
"2.up_proj": 0.04971493035554886,
"2.k_proj": 0.023553118109703064,
"2.o_proj": 0.020216671749949455,
"2.q_proj": 0.022627107799053192,
"2.v_proj": 0.0288323312997818,
"3.down_proj": 0.017859240993857384,
"3.gate_proj": 0.0461166575551033,
"3.up_proj": 0.04656068980693817,
"3.k_proj": 0.02325223572552204,
"3.o_proj": 0.01974094845354557,
"3.q_proj": 0.022899625822901726,
"3.v_proj": 0.0293908528983593,
"4.down_proj": 0.015974679961800575,
"4.gate_proj": 0.04187997430562973,
"4.up_proj": 0.04083123430609703,
"4.k_proj": 0.024421246722340584,
"4.o_proj": 0.01839449815452099,
"4.q_proj": 0.02253827266395092,
"4.v_proj": 0.026954350993037224,
"5.down_proj": 0.016185220330953598,
"5.gate_proj": 0.0418110266327858,
"5.up_proj": 0.040592070668935776,
"5.k_proj": 0.022455476224422455,
"5.o_proj": 0.017546743154525757,
"5.q_proj": 0.020589888095855713,
"5.v_proj": 0.026903508231043816,
"6.down_proj": 0.01673140563070774,
"6.gate_proj": 0.039437901228666306,
"6.up_proj": 0.037406571209430695,
"6.k_proj": 0.022581404075026512,
"6.o_proj": 0.017793400213122368,
"6.q_proj": 0.01947803795337677,
"6.v_proj": 0.025546329095959663,
"7.down_proj": 0.01703973300755024,
"7.gate_proj": 0.03845927119255066,
"7.up_proj": 0.042372316122055054,
"7.k_proj": 0.01995324343442917,
"7.o_proj": 0.01723671145737171,
"7.q_proj": 0.018467623740434647,
"7.v_proj": 0.025217900052666664,
"8.down_proj": 0.016596227884292603,
"8.gate_proj": 0.04214202240109444,
"8.up_proj": 0.03909769654273987,
"8.k_proj": 0.020271550863981247,
"8.o_proj": 0.016867313534021378,
"8.q_proj": 0.019599435850977898,
"8.v_proj": 0.02347864769399166,
"9.down_proj": 0.016085432842373848,
"9.gate_proj": 0.03925571218132973,
"9.up_proj": 0.03977390378713608,
"9.k_proj": 0.020905086770653725,
"9.o_proj": 0.017597880214452744,
"9.q_proj": 0.019629141315817833,
"9.v_proj": 0.021614503115415573,
"10.down_proj": 0.016450710594654083,
"10.gate_proj": 0.0410277396440506,
"10.up_proj": 0.036953117698431015,
"10.k_proj": 0.021330196410417557,
"10.o_proj": 0.018903516232967377,
"10.q_proj": 0.01969883218407631,
"10.v_proj": 0.02067733369767666,
"11.down_proj": 0.01707199402153492,
"11.gate_proj": 0.038356050848960876,
"11.up_proj": 0.041197046637535095,
"11.k_proj": 0.023145737126469612,
"11.o_proj": 0.016748346388339996,
"11.q_proj": 0.021463047713041306,
"11.v_proj": 0.019644424319267273,
"12.down_proj": 0.0163368321955204,
"12.gate_proj": 0.03845346346497536,
"12.up_proj": 0.039496049284935,
"12.k_proj": 0.022679366171360016,
"12.o_proj": 0.017901107668876648,
"12.q_proj": 0.021295785903930664,
"12.v_proj": 0.020925084128975868,
"13.down_proj": 0.016835162416100502,
"13.gate_proj": 0.04033529385924339,
"13.up_proj": 0.04001957178115845,
"13.k_proj": 0.02491459809243679,
"13.o_proj": 0.018841315060853958,
"13.q_proj": 0.022686036303639412,
"13.v_proj": 0.022567175328731537,
"14.down_proj": 0.01779063418507576,
"14.gate_proj": 0.04157009720802307,
"14.up_proj": 0.041127171367406845,
"14.k_proj": 0.025756051763892174,
"14.o_proj": 0.01769130863249302,
"14.q_proj": 0.024447226896882057,
"14.v_proj": 0.02386847883462906,
"15.down_proj": 0.018620144575834274,
"15.gate_proj": 0.04322356358170509,
"15.up_proj": 0.04224470630288124,
"15.k_proj": 0.025942062959074974,
"15.o_proj": 0.018257489427924156,
"15.q_proj": 0.023428166285157204,
"15.v_proj": 0.022524960339069366,
"16.down_proj": 0.018772562965750694,
"16.gate_proj": 0.044220734387636185,
"16.up_proj": 0.04224252700805664,
"16.k_proj": 0.029059557244181633,
"16.o_proj": 0.019190525636076927,
"16.q_proj": 0.025546090677380562,
"16.v_proj": 0.027710048481822014,
"17.down_proj": 0.019686385989189148,
"17.gate_proj": 0.04135333001613617,
"17.up_proj": 0.04486946761608124,
"17.k_proj": 0.02745143696665764,
"17.o_proj": 0.020055420696735382,
"17.q_proj": 0.025585098192095757,
"17.v_proj": 0.02658572606742382,
"18.down_proj": 0.020166408270597458,
"18.gate_proj": 0.041446503251791,
"18.up_proj": 0.04579155147075653,
"18.k_proj": 0.02984543703496456,
"18.o_proj": 0.019872967153787613,
"18.q_proj": 0.02715788222849369,
"18.v_proj": 0.029887434095144272,
"19.down_proj": 0.020344026386737823,
"19.gate_proj": 0.04185543581843376,
"19.up_proj": 0.047275569289922714,
"19.k_proj": 0.02909199893474579,
"19.o_proj": 0.022002629935741425,
"19.q_proj": 0.02898808941245079,
"19.v_proj": 0.028401562944054604,
"20.down_proj": 0.020384257659316063,
"20.gate_proj": 0.04890038073062897,
"20.up_proj": 0.04913109540939331,
"20.k_proj": 0.030232349410653114,
"20.o_proj": 0.022863484919071198,
"20.q_proj": 0.029142558574676514,
"20.v_proj": 0.030943676829338074,
"21.down_proj": 0.020678570494055748,
"21.gate_proj": 0.04660072550177574,
"21.up_proj": 0.04657881334424019,
"21.k_proj": 0.02797800861299038,
"21.o_proj": 0.024201812222599983,
"21.q_proj": 0.028927212581038475,
"21.v_proj": 0.032502662390470505,
"22.down_proj": 0.02047550119459629,
"22.gate_proj": 0.048371393233537674,
"22.up_proj": 0.053299009799957275,
"22.k_proj": 0.02902030199766159,
"22.o_proj": 0.021341780200600624,
"22.q_proj": 0.028871284797787666,
"22.v_proj": 0.034008994698524475,
"23.down_proj": 0.02101362682878971,
"23.gate_proj": 0.04909517243504524,
"23.up_proj": 0.05019266530871391,
"23.k_proj": 0.02889767661690712,
"23.o_proj": 0.02051742933690548,
"23.q_proj": 0.02883676066994667,
"23.v_proj": 0.03221878036856651,
"24.down_proj": 0.020756598562002182,
"24.gate_proj": 0.049081768840551376,
"24.up_proj": 0.05525779724121094,
"24.k_proj": 0.028636103495955467,
"24.o_proj": 0.02002854458987713,
"24.q_proj": 0.028798028826713562,
"24.v_proj": 0.03184721618890762,
"25.down_proj": 0.019914736971259117,
"25.gate_proj": 0.05659620091319084,
"25.up_proj": 0.05105341598391533,
"25.k_proj": 0.03020135685801506,
"25.o_proj": 0.02426576428115368,
"25.q_proj": 0.028738798573613167,
"25.v_proj": 0.030420735478401184,
"26.down_proj": 0.020024362951517105,
"26.gate_proj": 0.049845531582832336,
"26.up_proj": 0.047202229499816895,
"26.k_proj": 0.03258863091468811,
"26.o_proj": 0.01930871419608593,
"26.q_proj": 0.028618991374969482,
"26.v_proj": 0.030448047444224358,
"27.down_proj": 0.01905849389731884,
"27.gate_proj": 0.04903184995055199,
"27.up_proj": 0.0499153696000576,
"27.k_proj": 0.027926409617066383,
"27.o_proj": 0.02700573392212391,
"27.q_proj": 0.02987619861960411,
"27.v_proj": 0.0307757705450058,
"28.down_proj": 0.02064599096775055,
"28.gate_proj": 0.05176892504096031,
"28.up_proj": 0.04868489131331444,
"28.k_proj": 0.030534468591213226,
"28.o_proj": 0.01911034807562828,
"28.q_proj": 0.03064342774450779,
"28.v_proj": 0.027126865461468697,
"29.down_proj": 0.01944028027355671,
"29.gate_proj": 0.05063900351524353,
"29.up_proj": 0.04600493237376213,
"29.k_proj": 0.034188102930784225,
"29.o_proj": 0.01887245662510395,
"29.q_proj": 0.02965513803064823,
"29.v_proj": 0.03125184401869774,
"30.down_proj": 0.014209754765033722,
"30.gate_proj": 0.04668048024177551,
"30.up_proj": 0.043189410120248795,
"30.k_proj": 0.031168438494205475,
"30.o_proj": 0.018736323341727257,
"30.q_proj": 0.03064728155732155,
"30.v_proj": 0.022416142746806145,
"31.down_proj": 0.01456467341631651,
"31.gate_proj": 0.048269350081682205,
"31.up_proj": 0.04480603337287903,
"31.k_proj": 0.03716890141367912,
"31.o_proj": 0.018774691969156265,
"31.q_proj": 0.034700050950050354,
"31.v_proj": 0.02298746630549431
}
}