svd-safety-l2_basis_remove50 / compression.json
Jeesup's picture
None arm at None budget
78a8bc8 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": 0.5,
"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_remove50/truncated/factored.pt",
"param_fraction": 0.49982796308290156,
"ranks": {
"0.q_proj": 1365,
"0.k_proj": 1365,
"0.v_proj": 1365,
"0.o_proj": 1024,
"0.gate_proj": 1726,
"0.up_proj": 1726,
"0.down_proj": 1492,
"1.q_proj": 1365,
"1.k_proj": 1365,
"1.v_proj": 1365,
"1.o_proj": 1024,
"1.gate_proj": 1726,
"1.up_proj": 1726,
"1.down_proj": 1492,
"2.q_proj": 1365,
"2.k_proj": 1365,
"2.v_proj": 1365,
"2.o_proj": 1024,
"2.gate_proj": 1726,
"2.up_proj": 1726,
"2.down_proj": 1492,
"3.q_proj": 1365,
"3.k_proj": 1365,
"3.v_proj": 1365,
"3.o_proj": 1024,
"3.gate_proj": 1726,
"3.up_proj": 1726,
"3.down_proj": 1492,
"4.q_proj": 1365,
"4.k_proj": 1365,
"4.v_proj": 1365,
"4.o_proj": 1024,
"4.gate_proj": 1726,
"4.up_proj": 1726,
"4.down_proj": 1492,
"5.q_proj": 1365,
"5.k_proj": 1365,
"5.v_proj": 1365,
"5.o_proj": 1024,
"5.gate_proj": 1726,
"5.up_proj": 1726,
"5.down_proj": 1492,
"6.q_proj": 1365,
"6.k_proj": 1365,
"6.v_proj": 1365,
"6.o_proj": 1024,
"6.gate_proj": 1726,
"6.up_proj": 1726,
"6.down_proj": 1492,
"7.q_proj": 1365,
"7.k_proj": 1365,
"7.v_proj": 1365,
"7.o_proj": 1024,
"7.gate_proj": 1726,
"7.up_proj": 1726,
"7.down_proj": 1492,
"8.q_proj": 1365,
"8.k_proj": 1365,
"8.v_proj": 1365,
"8.o_proj": 1024,
"8.gate_proj": 1726,
"8.up_proj": 1726,
"8.down_proj": 1492,
"9.q_proj": 1365,
"9.k_proj": 1365,
"9.v_proj": 1365,
"9.o_proj": 1024,
"9.gate_proj": 1726,
"9.up_proj": 1726,
"9.down_proj": 1492,
"10.q_proj": 1365,
"10.k_proj": 1365,
"10.v_proj": 1365,
"10.o_proj": 1024,
"10.gate_proj": 1726,
"10.up_proj": 1726,
"10.down_proj": 1492,
"11.q_proj": 1365,
"11.k_proj": 1365,
"11.v_proj": 1365,
"11.o_proj": 1024,
"11.gate_proj": 1726,
"11.up_proj": 1726,
"11.down_proj": 1492,
"12.q_proj": 1365,
"12.k_proj": 1365,
"12.v_proj": 1365,
"12.o_proj": 1024,
"12.gate_proj": 1726,
"12.up_proj": 1726,
"12.down_proj": 1492,
"13.q_proj": 1365,
"13.k_proj": 1365,
"13.v_proj": 1365,
"13.o_proj": 1024,
"13.gate_proj": 1726,
"13.up_proj": 1726,
"13.down_proj": 1492,
"14.q_proj": 1365,
"14.k_proj": 1365,
"14.v_proj": 1365,
"14.o_proj": 1024,
"14.gate_proj": 1726,
"14.up_proj": 1726,
"14.down_proj": 1492,
"15.q_proj": 1365,
"15.k_proj": 1365,
"15.v_proj": 1365,
"15.o_proj": 1024,
"15.gate_proj": 1726,
"15.up_proj": 1726,
"15.down_proj": 1492,
"16.q_proj": 1365,
"16.k_proj": 1365,
"16.v_proj": 1365,
"16.o_proj": 1024,
"16.gate_proj": 1726,
"16.up_proj": 1726,
"16.down_proj": 1492,
"17.q_proj": 1365,
"17.k_proj": 1365,
"17.v_proj": 1365,
"17.o_proj": 1024,
"17.gate_proj": 1726,
"17.up_proj": 1726,
"17.down_proj": 1492,
"18.q_proj": 1365,
"18.k_proj": 1365,
"18.v_proj": 1365,
"18.o_proj": 1024,
"18.gate_proj": 1726,
"18.up_proj": 1726,
"18.down_proj": 1492,
"19.q_proj": 1365,
"19.k_proj": 1365,
"19.v_proj": 1365,
"19.o_proj": 1024,
"19.gate_proj": 1726,
"19.up_proj": 1726,
"19.down_proj": 1492,
"20.q_proj": 1365,
"20.k_proj": 1365,
"20.v_proj": 1365,
"20.o_proj": 1024,
"20.gate_proj": 1726,
"20.up_proj": 1726,
"20.down_proj": 1492,
"21.q_proj": 1365,
"21.k_proj": 1365,
"21.v_proj": 1365,
"21.o_proj": 1024,
"21.gate_proj": 1726,
"21.up_proj": 1726,
"21.down_proj": 1492,
"22.q_proj": 1365,
"22.k_proj": 1365,
"22.v_proj": 1365,
"22.o_proj": 1024,
"22.gate_proj": 1726,
"22.up_proj": 1726,
"22.down_proj": 1492,
"23.q_proj": 1365,
"23.k_proj": 1365,
"23.v_proj": 1365,
"23.o_proj": 1024,
"23.gate_proj": 1726,
"23.up_proj": 1726,
"23.down_proj": 1492,
"24.q_proj": 1365,
"24.k_proj": 1365,
"24.v_proj": 1365,
"24.o_proj": 1024,
"24.gate_proj": 1726,
"24.up_proj": 1726,
"24.down_proj": 1492,
"25.q_proj": 1365,
"25.k_proj": 1365,
"25.v_proj": 1365,
"25.o_proj": 1024,
"25.gate_proj": 1726,
"25.up_proj": 1726,
"25.down_proj": 1492,
"26.q_proj": 1365,
"26.k_proj": 1365,
"26.v_proj": 1365,
"26.o_proj": 1024,
"26.gate_proj": 1726,
"26.up_proj": 1726,
"26.down_proj": 1492,
"27.q_proj": 1365,
"27.k_proj": 1365,
"27.v_proj": 1365,
"27.o_proj": 1024,
"27.gate_proj": 1726,
"27.up_proj": 1726,
"27.down_proj": 1492,
"28.q_proj": 1365,
"28.k_proj": 1365,
"28.v_proj": 1365,
"28.o_proj": 1024,
"28.gate_proj": 1726,
"28.up_proj": 1726,
"28.down_proj": 1492,
"29.q_proj": 1365,
"29.k_proj": 1365,
"29.v_proj": 1365,
"29.o_proj": 1024,
"29.gate_proj": 1726,
"29.up_proj": 1726,
"29.down_proj": 1492,
"30.q_proj": 1365,
"30.k_proj": 1365,
"30.v_proj": 1365,
"30.o_proj": 1024,
"30.gate_proj": 1726,
"30.up_proj": 1726,
"30.down_proj": 1492,
"31.q_proj": 1365,
"31.k_proj": 1365,
"31.v_proj": 1365,
"31.o_proj": 1024,
"31.gate_proj": 1726,
"31.up_proj": 1726,
"31.down_proj": 1492
},
"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": 0.03477951146695497
},
"param_fraction_is_upper_bound": false,
"rel_delta_coefficient": {
"0.down_proj": 0.023729722946882248,
"0.gate_proj": 0.08452913910150528,
"0.up_proj": 0.059868842363357544,
"0.k_proj": 0.05120539665222168,
"0.o_proj": 0.017369695007801056,
"0.q_proj": 0.056597448885440826,
"0.v_proj": 0.026993343606591225,
"1.down_proj": 0.01011702511459589,
"1.gate_proj": 0.06331334263086319,
"1.up_proj": 0.052683912217617035,
"1.k_proj": 0.02192487008869648,
"1.o_proj": 0.01861078478395939,
"1.q_proj": 0.027439886704087257,
"1.v_proj": 0.019045013934373856,
"2.down_proj": 0.0223842766135931,
"2.gate_proj": 0.060128744691610336,
"2.up_proj": 0.0586223378777504,
"2.k_proj": 0.026699945330619812,
"2.o_proj": 0.021796977147459984,
"2.q_proj": 0.028158212080597878,
"2.v_proj": 0.03134380280971527,
"3.down_proj": 0.022037558257579803,
"3.gate_proj": 0.053900573402643204,
"3.up_proj": 0.05270858481526375,
"3.k_proj": 0.02719547599554062,
"3.o_proj": 0.021815096959471703,
"3.q_proj": 0.026857580989599228,
"3.v_proj": 0.03770525008440018,
"4.down_proj": 0.02017192542552948,
"4.gate_proj": 0.04923078790307045,
"4.up_proj": 0.04876559600234032,
"4.k_proj": 0.03054051846265793,
"4.o_proj": 0.024025147780776024,
"4.q_proj": 0.026279738172888756,
"4.v_proj": 0.02896757237613201,
"5.down_proj": 0.019595874473452568,
"5.gate_proj": 0.04980846494436264,
"5.up_proj": 0.04777255654335022,
"5.k_proj": 0.02639787830412388,
"5.o_proj": 0.01787957362830639,
"5.q_proj": 0.021695686504244804,
"5.v_proj": 0.030705556273460388,
"6.down_proj": 0.01879432424902916,
"6.gate_proj": 0.045955441892147064,
"6.up_proj": 0.046904079616069794,
"6.k_proj": 0.025844892486929893,
"6.o_proj": 0.020820025354623795,
"6.q_proj": 0.0238942988216877,
"6.v_proj": 0.029350096359848976,
"7.down_proj": 0.020194197073578835,
"7.gate_proj": 0.04724954441189766,
"7.up_proj": 0.04511765390634537,
"7.k_proj": 0.025140712037682533,
"7.o_proj": 0.019103750586509705,
"7.q_proj": 0.0233695600181818,
"7.v_proj": 0.027190091088414192,
"8.down_proj": 0.01932426728308201,
"8.gate_proj": 0.04175572097301483,
"8.up_proj": 0.044960521161556244,
"8.k_proj": 0.02549305558204651,
"8.o_proj": 0.020664075389504433,
"8.q_proj": 0.022841913625597954,
"8.v_proj": 0.025211365893483162,
"9.down_proj": 0.017424633726477623,
"9.gate_proj": 0.047231525182724,
"9.up_proj": 0.04754003509879112,
"9.k_proj": 0.02622208558022976,
"9.o_proj": 0.018889963626861572,
"9.q_proj": 0.024236785247921944,
"9.v_proj": 0.02502570115029812,
"10.down_proj": 0.018354928120970726,
"10.gate_proj": 0.0459887869656086,
"10.up_proj": 0.04865645617246628,
"10.k_proj": 0.02619076706469059,
"10.o_proj": 0.021123865619301796,
"10.q_proj": 0.023364614695310593,
"10.v_proj": 0.024980315938591957,
"11.down_proj": 0.01799744740128517,
"11.gate_proj": 0.04674726724624634,
"11.up_proj": 0.04505652189254761,
"11.k_proj": 0.026833688840270042,
"11.o_proj": 0.01736118644475937,
"11.q_proj": 0.023478467017412186,
"11.v_proj": 0.019759302958846092,
"12.down_proj": 0.019656836986541748,
"12.gate_proj": 0.04380928352475166,
"12.up_proj": 0.046775877475738525,
"12.k_proj": 0.027490461245179176,
"12.o_proj": 0.019591068848967552,
"12.q_proj": 0.026375921443104744,
"12.v_proj": 0.02554561384022236,
"13.down_proj": 0.02064443565905094,
"13.gate_proj": 0.04611923545598984,
"13.up_proj": 0.049012213945388794,
"13.k_proj": 0.029619509354233742,
"13.o_proj": 0.0220456775277853,
"13.q_proj": 0.028390999883413315,
"13.v_proj": 0.027894970029592514,
"14.down_proj": 0.02147248387336731,
"14.gate_proj": 0.046727463603019714,
"14.up_proj": 0.04966231808066368,
"14.k_proj": 0.029388461261987686,
"14.o_proj": 0.02229248732328415,
"14.q_proj": 0.026943136006593704,
"14.v_proj": 0.027938341721892357,
"15.down_proj": 0.02149982564151287,
"15.gate_proj": 0.050257809460163116,
"15.up_proj": 0.04973958060145378,
"15.k_proj": 0.03036561980843544,
"15.o_proj": 0.022304357960820198,
"15.q_proj": 0.02832680568099022,
"15.v_proj": 0.026476964354515076,
"16.down_proj": 0.02218230813741684,
"16.gate_proj": 0.050966568291187286,
"16.up_proj": 0.05037204548716545,
"16.k_proj": 0.032374244183301926,
"16.o_proj": 0.022258751094341278,
"16.q_proj": 0.030399447306990623,
"16.v_proj": 0.029645616188645363,
"17.down_proj": 0.02408037707209587,
"17.gate_proj": 0.05045311152935028,
"17.up_proj": 0.052991095930337906,
"17.k_proj": 0.03171582520008087,
"17.o_proj": 0.024419406428933144,
"17.q_proj": 0.029634423553943634,
"17.v_proj": 0.029784325510263443,
"18.down_proj": 0.024372927844524384,
"18.gate_proj": 0.048495788127183914,
"18.up_proj": 0.05652078613638878,
"18.k_proj": 0.03411583974957466,
"18.o_proj": 0.025013010948896408,
"18.q_proj": 0.033185336738824844,
"18.v_proj": 0.033583350479602814,
"19.down_proj": 0.023557795211672783,
"19.gate_proj": 0.04933939501643181,
"19.up_proj": 0.05396072566509247,
"19.k_proj": 0.03440993279218674,
"19.o_proj": 0.026062030345201492,
"19.q_proj": 0.032321497797966,
"19.v_proj": 0.033995699137449265,
"20.down_proj": 0.024951640516519547,
"20.gate_proj": 0.04994823411107063,
"20.up_proj": 0.05292683094739914,
"20.k_proj": 0.03518199548125267,
"20.o_proj": 0.026832351461052895,
"20.q_proj": 0.03478262200951576,
"20.v_proj": 0.034277647733688354,
"21.down_proj": 0.024847429245710373,
"21.gate_proj": 0.05355115979909897,
"21.up_proj": 0.0550607293844223,
"21.k_proj": 0.03435465320944786,
"21.o_proj": 0.02734820917248726,
"21.q_proj": 0.03575967997312546,
"21.v_proj": 0.039932653307914734,
"22.down_proj": 0.025028206408023834,
"22.gate_proj": 0.05792507156729698,
"22.up_proj": 0.056496161967515945,
"22.k_proj": 0.035662177950143814,
"22.o_proj": 0.02528970316052437,
"22.q_proj": 0.03271043300628662,
"22.v_proj": 0.03736736625432968,
"23.down_proj": 0.024994004517793655,
"23.gate_proj": 0.05386468768119812,
"23.up_proj": 0.05626451596617699,
"23.k_proj": 0.03418954834342003,
"23.o_proj": 0.023994209244847298,
"23.q_proj": 0.03303827345371246,
"23.v_proj": 0.03754240646958351,
"24.down_proj": 0.025210509076714516,
"24.gate_proj": 0.05693400278687477,
"24.up_proj": 0.06670183688402176,
"24.k_proj": 0.03441649302840233,
"24.o_proj": 0.021066762506961823,
"24.q_proj": 0.0341021791100502,
"24.v_proj": 0.03781954199075699,
"25.down_proj": 0.02424655854701996,
"25.gate_proj": 0.06084345653653145,
"25.up_proj": 0.06915346533060074,
"25.k_proj": 0.035405270755290985,
"25.o_proj": 0.02484528347849846,
"25.q_proj": 0.033545415848493576,
"25.v_proj": 0.0366339273750782,
"26.down_proj": 0.0257209874689579,
"26.gate_proj": 0.05857868492603302,
"26.up_proj": 0.05918649584054947,
"26.k_proj": 0.04149233549833298,
"26.o_proj": 0.02095915749669075,
"26.q_proj": 0.036208391189575195,
"26.v_proj": 0.039256542921066284,
"27.down_proj": 0.02435978874564171,
"27.gate_proj": 0.057182516902685165,
"27.up_proj": 0.05642085522413254,
"27.k_proj": 0.03608841821551323,
"27.o_proj": 0.031146977096796036,
"27.q_proj": 0.035692207515239716,
"27.v_proj": 0.036948930472135544,
"28.down_proj": 0.0241420716047287,
"28.gate_proj": 0.058406781405210495,
"28.up_proj": 0.05536749213933945,
"28.k_proj": 0.037475209683179855,
"28.o_proj": 0.025948751717805862,
"28.q_proj": 0.03311076760292053,
"28.v_proj": 0.036046117544174194,
"29.down_proj": 0.023028584197163582,
"29.gate_proj": 0.05505272001028061,
"29.up_proj": 0.058274224400520325,
"29.k_proj": 0.04002455994486809,
"29.o_proj": 0.021812763065099716,
"29.q_proj": 0.03628501296043396,
"29.v_proj": 0.0329945832490921,
"30.down_proj": 0.014910749159753323,
"30.gate_proj": 0.0535011924803257,
"30.up_proj": 0.05174972116947174,
"30.k_proj": 0.038816239684820175,
"30.o_proj": 0.023247702047228813,
"30.q_proj": 0.034509968012571335,
"30.v_proj": 0.023940032348036766,
"31.down_proj": 0.01771707274019718,
"31.gate_proj": 0.0577886700630188,
"31.up_proj": 0.04945879802107811,
"31.k_proj": 0.04337715357542038,
"31.o_proj": 0.019638074561953545,
"31.q_proj": 0.039459045976400375,
"31.v_proj": 0.025393839925527573
}
}