svd-safety-l2_basis_remove30 / compression.json
Jeesup's picture
None arm at None budget
9f6f4bb 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": 0.3,
"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_remove30/truncated/factored.pt",
"param_fraction": 0.6998272040965026,
"ranks": {
"0.q_proj": 1911,
"0.k_proj": 1911,
"0.v_proj": 1911,
"0.o_proj": 1433,
"0.gate_proj": 2417,
"0.up_proj": 2417,
"0.down_proj": 2089,
"1.q_proj": 1911,
"1.k_proj": 1911,
"1.v_proj": 1911,
"1.o_proj": 1433,
"1.gate_proj": 2417,
"1.up_proj": 2417,
"1.down_proj": 2089,
"2.q_proj": 1911,
"2.k_proj": 1911,
"2.v_proj": 1911,
"2.o_proj": 1433,
"2.gate_proj": 2417,
"2.up_proj": 2417,
"2.down_proj": 2089,
"3.q_proj": 1911,
"3.k_proj": 1911,
"3.v_proj": 1911,
"3.o_proj": 1433,
"3.gate_proj": 2417,
"3.up_proj": 2417,
"3.down_proj": 2089,
"4.q_proj": 1911,
"4.k_proj": 1911,
"4.v_proj": 1911,
"4.o_proj": 1433,
"4.gate_proj": 2417,
"4.up_proj": 2417,
"4.down_proj": 2089,
"5.q_proj": 1911,
"5.k_proj": 1911,
"5.v_proj": 1911,
"5.o_proj": 1433,
"5.gate_proj": 2417,
"5.up_proj": 2417,
"5.down_proj": 2089,
"6.q_proj": 1911,
"6.k_proj": 1911,
"6.v_proj": 1911,
"6.o_proj": 1433,
"6.gate_proj": 2417,
"6.up_proj": 2417,
"6.down_proj": 2089,
"7.q_proj": 1911,
"7.k_proj": 1911,
"7.v_proj": 1911,
"7.o_proj": 1433,
"7.gate_proj": 2417,
"7.up_proj": 2417,
"7.down_proj": 2089,
"8.q_proj": 1911,
"8.k_proj": 1911,
"8.v_proj": 1911,
"8.o_proj": 1433,
"8.gate_proj": 2417,
"8.up_proj": 2417,
"8.down_proj": 2089,
"9.q_proj": 1911,
"9.k_proj": 1911,
"9.v_proj": 1911,
"9.o_proj": 1433,
"9.gate_proj": 2417,
"9.up_proj": 2417,
"9.down_proj": 2089,
"10.q_proj": 1911,
"10.k_proj": 1911,
"10.v_proj": 1911,
"10.o_proj": 1433,
"10.gate_proj": 2417,
"10.up_proj": 2417,
"10.down_proj": 2089,
"11.q_proj": 1911,
"11.k_proj": 1911,
"11.v_proj": 1911,
"11.o_proj": 1433,
"11.gate_proj": 2417,
"11.up_proj": 2417,
"11.down_proj": 2089,
"12.q_proj": 1911,
"12.k_proj": 1911,
"12.v_proj": 1911,
"12.o_proj": 1433,
"12.gate_proj": 2417,
"12.up_proj": 2417,
"12.down_proj": 2089,
"13.q_proj": 1911,
"13.k_proj": 1911,
"13.v_proj": 1911,
"13.o_proj": 1433,
"13.gate_proj": 2417,
"13.up_proj": 2417,
"13.down_proj": 2089,
"14.q_proj": 1911,
"14.k_proj": 1911,
"14.v_proj": 1911,
"14.o_proj": 1433,
"14.gate_proj": 2417,
"14.up_proj": 2417,
"14.down_proj": 2089,
"15.q_proj": 1911,
"15.k_proj": 1911,
"15.v_proj": 1911,
"15.o_proj": 1433,
"15.gate_proj": 2417,
"15.up_proj": 2417,
"15.down_proj": 2089,
"16.q_proj": 1911,
"16.k_proj": 1911,
"16.v_proj": 1911,
"16.o_proj": 1433,
"16.gate_proj": 2417,
"16.up_proj": 2417,
"16.down_proj": 2089,
"17.q_proj": 1911,
"17.k_proj": 1911,
"17.v_proj": 1911,
"17.o_proj": 1433,
"17.gate_proj": 2417,
"17.up_proj": 2417,
"17.down_proj": 2089,
"18.q_proj": 1911,
"18.k_proj": 1911,
"18.v_proj": 1911,
"18.o_proj": 1433,
"18.gate_proj": 2417,
"18.up_proj": 2417,
"18.down_proj": 2089,
"19.q_proj": 1911,
"19.k_proj": 1911,
"19.v_proj": 1911,
"19.o_proj": 1433,
"19.gate_proj": 2417,
"19.up_proj": 2417,
"19.down_proj": 2089,
"20.q_proj": 1911,
"20.k_proj": 1911,
"20.v_proj": 1911,
"20.o_proj": 1433,
"20.gate_proj": 2417,
"20.up_proj": 2417,
"20.down_proj": 2089,
"21.q_proj": 1911,
"21.k_proj": 1911,
"21.v_proj": 1911,
"21.o_proj": 1433,
"21.gate_proj": 2417,
"21.up_proj": 2417,
"21.down_proj": 2089,
"22.q_proj": 1911,
"22.k_proj": 1911,
"22.v_proj": 1911,
"22.o_proj": 1433,
"22.gate_proj": 2417,
"22.up_proj": 2417,
"22.down_proj": 2089,
"23.q_proj": 1911,
"23.k_proj": 1911,
"23.v_proj": 1911,
"23.o_proj": 1433,
"23.gate_proj": 2417,
"23.up_proj": 2417,
"23.down_proj": 2089,
"24.q_proj": 1911,
"24.k_proj": 1911,
"24.v_proj": 1911,
"24.o_proj": 1433,
"24.gate_proj": 2417,
"24.up_proj": 2417,
"24.down_proj": 2089,
"25.q_proj": 1911,
"25.k_proj": 1911,
"25.v_proj": 1911,
"25.o_proj": 1433,
"25.gate_proj": 2417,
"25.up_proj": 2417,
"25.down_proj": 2089,
"26.q_proj": 1911,
"26.k_proj": 1911,
"26.v_proj": 1911,
"26.o_proj": 1433,
"26.gate_proj": 2417,
"26.up_proj": 2417,
"26.down_proj": 2089,
"27.q_proj": 1911,
"27.k_proj": 1911,
"27.v_proj": 1911,
"27.o_proj": 1433,
"27.gate_proj": 2417,
"27.up_proj": 2417,
"27.down_proj": 2089,
"28.q_proj": 1911,
"28.k_proj": 1911,
"28.v_proj": 1911,
"28.o_proj": 1433,
"28.gate_proj": 2417,
"28.up_proj": 2417,
"28.down_proj": 2089,
"29.q_proj": 1911,
"29.k_proj": 1911,
"29.v_proj": 1911,
"29.o_proj": 1433,
"29.gate_proj": 2417,
"29.up_proj": 2417,
"29.down_proj": 2089,
"30.q_proj": 1911,
"30.k_proj": 1911,
"30.v_proj": 1911,
"30.o_proj": 1433,
"30.gate_proj": 2417,
"30.up_proj": 2417,
"30.down_proj": 2089,
"31.q_proj": 1911,
"31.k_proj": 1911,
"31.v_proj": 1911,
"31.o_proj": 1433,
"31.gate_proj": 2417,
"31.up_proj": 2417,
"31.down_proj": 2089
},
"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": 0.026299988264716894
},
"param_fraction_is_upper_bound": false,
"rel_delta_coefficient": {
"0.down_proj": 0.01571832410991192,
"0.gate_proj": 0.07073784619569778,
"0.up_proj": 0.051439255475997925,
"0.k_proj": 0.03125172480940819,
"0.o_proj": 0.014449120499193668,
"0.q_proj": 0.03769339993596077,
"0.v_proj": 0.019454704597592354,
"1.down_proj": 0.007780988700687885,
"1.gate_proj": 0.05042870715260506,
"1.up_proj": 0.04310716688632965,
"1.k_proj": 0.01711668074131012,
"1.o_proj": 0.013755874708294868,
"1.q_proj": 0.01987244002521038,
"1.v_proj": 0.015976836904883385,
"2.down_proj": 0.01568828709423542,
"2.gate_proj": 0.04502059519290924,
"2.up_proj": 0.04337015375494957,
"2.k_proj": 0.01923982799053192,
"2.o_proj": 0.018201353028416634,
"2.q_proj": 0.019387632608413696,
"2.v_proj": 0.025767281651496887,
"3.down_proj": 0.015289364382624626,
"3.gate_proj": 0.03861368075013161,
"3.up_proj": 0.040375787764787674,
"3.k_proj": 0.019376985728740692,
"3.o_proj": 0.017843443900346756,
"3.q_proj": 0.01794094406068325,
"3.v_proj": 0.025516241788864136,
"4.down_proj": 0.013919559307396412,
"4.gate_proj": 0.03799600154161453,
"4.up_proj": 0.03619149327278137,
"4.k_proj": 0.021520331501960754,
"4.o_proj": 0.017562542110681534,
"4.q_proj": 0.019273044541478157,
"4.v_proj": 0.02358822524547577,
"5.down_proj": 0.014357398264110088,
"5.gate_proj": 0.03674551099538803,
"5.up_proj": 0.033642545342445374,
"5.k_proj": 0.019489075988531113,
"5.o_proj": 0.01640571840107441,
"5.q_proj": 0.017913563176989555,
"5.v_proj": 0.02251928113400936,
"6.down_proj": 0.013552287593483925,
"6.gate_proj": 0.037070877850055695,
"6.up_proj": 0.03368235006928444,
"6.k_proj": 0.018359050154685974,
"6.o_proj": 0.01698077842593193,
"6.q_proj": 0.018099360167980194,
"6.v_proj": 0.024023892357945442,
"7.down_proj": 0.014833086170256138,
"7.gate_proj": 0.03736833110451698,
"7.up_proj": 0.03326217830181122,
"7.k_proj": 0.01821022666990757,
"7.o_proj": 0.014988509006798267,
"7.q_proj": 0.017297787591814995,
"7.v_proj": 0.022407837212085724,
"8.down_proj": 0.014573442749679089,
"8.gate_proj": 0.03457353264093399,
"8.up_proj": 0.033223461359739304,
"8.k_proj": 0.01891668699681759,
"8.o_proj": 0.016290225088596344,
"8.q_proj": 0.018463721498847008,
"8.v_proj": 0.020800800994038582,
"9.down_proj": 0.013626599684357643,
"9.gate_proj": 0.03514138609170914,
"9.up_proj": 0.03454797342419624,
"9.k_proj": 0.01954204961657524,
"9.o_proj": 0.01583862490952015,
"9.q_proj": 0.018615638837218285,
"9.v_proj": 0.020167481154203415,
"10.down_proj": 0.014242639765143394,
"10.gate_proj": 0.036446455866098404,
"10.up_proj": 0.03395068645477295,
"10.k_proj": 0.019043728709220886,
"10.o_proj": 0.016665460541844368,
"10.q_proj": 0.018057456240057945,
"10.v_proj": 0.019024686887860298,
"11.down_proj": 0.015268338844180107,
"11.gate_proj": 0.034873563796281815,
"11.up_proj": 0.03515440225601196,
"11.k_proj": 0.020495600998401642,
"11.o_proj": 0.01653156615793705,
"11.q_proj": 0.01963174156844616,
"11.v_proj": 0.018966326490044594,
"12.down_proj": 0.01490208599716425,
"12.gate_proj": 0.036391180008649826,
"12.up_proj": 0.03552509844303131,
"12.k_proj": 0.020210448652505875,
"12.o_proj": 0.015827499330043793,
"12.q_proj": 0.01948561705648899,
"12.v_proj": 0.020654533058404922,
"13.down_proj": 0.01405719481408596,
"13.gate_proj": 0.03551631048321724,
"13.up_proj": 0.037898264825344086,
"13.k_proj": 0.020987654104828835,
"13.o_proj": 0.016064412891864777,
"13.q_proj": 0.019807111471891403,
"13.v_proj": 0.01962696947157383,
"14.down_proj": 0.016009340062737465,
"14.gate_proj": 0.03802134096622467,
"14.up_proj": 0.03786147013306618,
"14.k_proj": 0.021262750029563904,
"14.o_proj": 0.017347760498523712,
"14.q_proj": 0.020732464268803596,
"14.v_proj": 0.02084069512784481,
"15.down_proj": 0.016418438404798508,
"15.gate_proj": 0.038650088012218475,
"15.up_proj": 0.035394005477428436,
"15.k_proj": 0.022153930738568306,
"15.o_proj": 0.01559238601475954,
"15.q_proj": 0.02090977504849434,
"15.v_proj": 0.02013670839369297,
"16.down_proj": 0.016190309077501297,
"16.gate_proj": 0.040097519755363464,
"16.up_proj": 0.03531859815120697,
"16.k_proj": 0.024484077468514442,
"16.o_proj": 0.018780086189508438,
"16.q_proj": 0.02221532352268696,
"16.v_proj": 0.02362828142940998,
"17.down_proj": 0.017163388431072235,
"17.gate_proj": 0.039196014404296875,
"17.up_proj": 0.03838655725121498,
"17.k_proj": 0.023485826328396797,
"17.o_proj": 0.01819983869791031,
"17.q_proj": 0.024406565353274345,
"17.v_proj": 0.023620449006557465,
"18.down_proj": 0.017160629853606224,
"18.gate_proj": 0.038988612592220306,
"18.up_proj": 0.04259044677019119,
"18.k_proj": 0.024072891101241112,
"18.o_proj": 0.019031178206205368,
"18.q_proj": 0.024145307019352913,
"18.v_proj": 0.02690170519053936,
"19.down_proj": 0.01757754199206829,
"19.gate_proj": 0.04096819460391998,
"19.up_proj": 0.03828761354088783,
"19.k_proj": 0.025333957746624947,
"19.o_proj": 0.01932433247566223,
"19.q_proj": 0.024755822494626045,
"19.v_proj": 0.02619953081011772,
"20.down_proj": 0.017618484795093536,
"20.gate_proj": 0.04168369621038437,
"20.up_proj": 0.04031258076429367,
"20.k_proj": 0.025555750355124474,
"20.o_proj": 0.019319145008921623,
"20.q_proj": 0.026066727936267853,
"20.v_proj": 0.0259451512247324,
"21.down_proj": 0.018069017678499222,
"21.gate_proj": 0.03980571776628494,
"21.up_proj": 0.04347342625260353,
"21.k_proj": 0.026453960686922073,
"21.o_proj": 0.01866300217807293,
"21.q_proj": 0.025375496596097946,
"21.v_proj": 0.027465440332889557,
"22.down_proj": 0.01788206212222576,
"22.gate_proj": 0.04741738364100456,
"22.up_proj": 0.04643930122256279,
"22.k_proj": 0.026044832542538643,
"22.o_proj": 0.018997332081198692,
"22.q_proj": 0.025562603026628494,
"22.v_proj": 0.029967913404107094,
"23.down_proj": 0.018382783979177475,
"23.gate_proj": 0.04516542702913284,
"23.up_proj": 0.04480283334851265,
"23.k_proj": 0.025396348908543587,
"23.o_proj": 0.016814321279525757,
"23.q_proj": 0.025049729272723198,
"23.v_proj": 0.0272342711687088,
"24.down_proj": 0.018060794100165367,
"24.gate_proj": 0.045451823621988297,
"24.up_proj": 0.04517007991671562,
"24.k_proj": 0.024571320042014122,
"24.o_proj": 0.015839651226997375,
"24.q_proj": 0.02506706677377224,
"24.v_proj": 0.030967289581894875,
"25.down_proj": 0.017680563032627106,
"25.gate_proj": 0.04438598453998566,
"25.up_proj": 0.04630052670836449,
"25.k_proj": 0.026787467300891876,
"25.o_proj": 0.018963590264320374,
"25.q_proj": 0.024902934208512306,
"25.v_proj": 0.02655676379799843,
"26.down_proj": 0.01728445664048195,
"26.gate_proj": 0.04320848360657692,
"26.up_proj": 0.04594198614358902,
"26.k_proj": 0.029789509251713753,
"26.o_proj": 0.01576789654791355,
"26.q_proj": 0.02549917809665203,
"26.v_proj": 0.027597865089774132,
"27.down_proj": 0.016805997118353844,
"27.gate_proj": 0.04197928309440613,
"27.up_proj": 0.04444769397377968,
"27.k_proj": 0.024795103818178177,
"27.o_proj": 0.023305708542466164,
"27.q_proj": 0.026696208864450455,
"27.v_proj": 0.02590230293571949,
"28.down_proj": 0.01710508018732071,
"28.gate_proj": 0.04344693571329117,
"28.up_proj": 0.041666075587272644,
"28.k_proj": 0.02632494829595089,
"28.o_proj": 0.017413025721907616,
"28.q_proj": 0.023984916508197784,
"28.v_proj": 0.029418673366308212,
"29.down_proj": 0.01671851985156536,
"29.gate_proj": 0.0447494238615036,
"29.up_proj": 0.04233136773109436,
"29.k_proj": 0.028628991916775703,
"29.o_proj": 0.01745348423719406,
"29.q_proj": 0.02534261718392372,
"29.v_proj": 0.026935601606965065,
"30.down_proj": 0.013234722428023815,
"30.gate_proj": 0.04196500778198242,
"30.up_proj": 0.0402400940656662,
"30.k_proj": 0.029129525646567345,
"30.o_proj": 0.018144818022847176,
"30.q_proj": 0.02681029587984085,
"30.v_proj": 0.019666632637381554,
"31.down_proj": 0.012836103327572346,
"31.gate_proj": 0.04107679799199104,
"31.up_proj": 0.038529664278030396,
"31.k_proj": 0.033828120678663254,
"31.o_proj": 0.016145970672369003,
"31.q_proj": 0.03200652077794075,
"31.v_proj": 0.022225258871912956
}
}