{ "source": "/workspace/svd_safety/assets/models/Mistral-7B-Instruct-v0.2", "upstream": "hiahei/Swift-SVD (compression) + AIoT-MLSys-Lab/SVD-LLM utils/LoRA.py (recovery)", "stage": "finetuned", "ratio_removed": 0.5, "internal_ratio_kept": 0.5, "from_pt": "/workspace/svd_safety/results/swift_ft/cells/mis7_swift_remove50/finetune_v/merge.pt", "mean_rel_err": 0.5888538884464651, "max_rel_err": 0.7068923115730286, "param_fraction": 0.5002723106971154, "ranks": { "0.q_proj": 1152, "0.k_proj": 485, "0.v_proj": 442, "0.o_proj": 1052, "0.gate_proj": 1680, "0.up_proj": 1680, "0.down_proj": 1702, "1.q_proj": 1157, "1.k_proj": 493, "1.v_proj": 465, "1.o_proj": 1084, "1.gate_proj": 1680, "1.up_proj": 1680, "1.down_proj": 1976, "2.q_proj": 942, "2.k_proj": 374, "2.v_proj": 367, "2.o_proj": 907, "2.gate_proj": 1680, "2.up_proj": 1680, "2.down_proj": 1351, "3.q_proj": 940, "3.k_proj": 372, "3.v_proj": 363, "3.o_proj": 898, "3.gate_proj": 1680, "3.up_proj": 1680, "3.down_proj": 1353, "4.q_proj": 938, "4.k_proj": 371, "4.v_proj": 366, "4.o_proj": 903, "4.gate_proj": 1680, "4.up_proj": 1680, "4.down_proj": 1371, "5.q_proj": 938, "5.k_proj": 372, "5.v_proj": 366, "5.o_proj": 910, "5.gate_proj": 1680, "5.up_proj": 1680, "5.down_proj": 1384, "6.q_proj": 939, "6.k_proj": 373, "6.v_proj": 368, "6.o_proj": 917, "6.gate_proj": 1680, "6.up_proj": 1680, "6.down_proj": 1395, "7.q_proj": 938, "7.k_proj": 373, "7.v_proj": 368, "7.o_proj": 925, "7.gate_proj": 1680, "7.up_proj": 1680, "7.down_proj": 1406, "8.q_proj": 935, "8.k_proj": 371, "8.v_proj": 368, "8.o_proj": 924, "8.gate_proj": 1680, "8.up_proj": 1680, "8.down_proj": 1403, "9.q_proj": 934, "9.k_proj": 370, "9.v_proj": 369, "9.o_proj": 924, "9.gate_proj": 1680, "9.up_proj": 1680, "9.down_proj": 1409, "10.q_proj": 935, "10.k_proj": 371, "10.v_proj": 369, "10.o_proj": 932, "10.gate_proj": 1680, "10.up_proj": 1680, "10.down_proj": 1410, "11.q_proj": 933, "11.k_proj": 370, "11.v_proj": 370, "11.o_proj": 934, "11.gate_proj": 1680, "11.up_proj": 1680, "11.down_proj": 1412, "12.q_proj": 939, "12.k_proj": 374, "12.v_proj": 373, "12.o_proj": 941, "12.gate_proj": 1680, "12.up_proj": 1680, "12.down_proj": 1428, "13.q_proj": 936, "13.k_proj": 372, "13.v_proj": 372, "13.o_proj": 939, "13.gate_proj": 1680, "13.up_proj": 1680, "13.down_proj": 1428, "14.q_proj": 944, "14.k_proj": 376, "14.v_proj": 378, "14.o_proj": 949, "14.gate_proj": 1680, "14.up_proj": 1680, "14.down_proj": 1451, "15.q_proj": 955, "15.k_proj": 383, "15.v_proj": 385, "15.o_proj": 964, "15.gate_proj": 1680, "15.up_proj": 1680, "15.down_proj": 1481, "16.q_proj": 959, "16.k_proj": 385, "16.v_proj": 387, "16.o_proj": 968, "16.gate_proj": 1680, "16.up_proj": 1680, "16.down_proj": 1500, "17.q_proj": 978, "17.k_proj": 395, "17.v_proj": 397, "17.o_proj": 990, "17.gate_proj": 1680, "17.up_proj": 1680, "17.down_proj": 1543, "18.q_proj": 978, "18.k_proj": 394, "18.v_proj": 398, "18.o_proj": 993, "18.gate_proj": 1680, "18.up_proj": 1680, "18.down_proj": 1555, "19.q_proj": 965, "19.k_proj": 388, "19.v_proj": 391, "19.o_proj": 983, "19.gate_proj": 1680, "19.up_proj": 1680, "19.down_proj": 1537, "20.q_proj": 942, "20.k_proj": 375, "20.v_proj": 379, "20.o_proj": 955, "20.gate_proj": 1680, "20.up_proj": 1680, "20.down_proj": 1488, "21.q_proj": 936, "21.k_proj": 371, "21.v_proj": 376, "21.o_proj": 947, "21.gate_proj": 1680, "21.up_proj": 1680, "21.down_proj": 1477, "22.q_proj": 926, "22.k_proj": 365, "22.v_proj": 371, "22.o_proj": 934, "22.gate_proj": 1680, "22.up_proj": 1680, "22.down_proj": 1457, "23.q_proj": 927, "23.k_proj": 365, "23.v_proj": 371, "23.o_proj": 940, "23.gate_proj": 1680, "23.up_proj": 1680, "23.down_proj": 1460, "24.q_proj": 926, "24.k_proj": 365, "24.v_proj": 371, "24.o_proj": 944, "24.gate_proj": 1680, "24.up_proj": 1680, "24.down_proj": 1461, "25.q_proj": 929, "25.k_proj": 366, "25.v_proj": 374, "25.o_proj": 951, "25.gate_proj": 1680, "25.up_proj": 1680, "25.down_proj": 1471, "26.q_proj": 933, "26.k_proj": 368, "26.v_proj": 377, "26.o_proj": 957, "26.gate_proj": 1680, "26.up_proj": 1680, "26.down_proj": 1483, "27.q_proj": 936, "27.k_proj": 370, "27.v_proj": 379, "27.o_proj": 958, "27.gate_proj": 1680, "27.up_proj": 1680, "27.down_proj": 1493, "28.q_proj": 949, "28.k_proj": 378, "28.v_proj": 387, "28.o_proj": 976, "28.gate_proj": 1680, "28.up_proj": 1680, "28.down_proj": 1532, "29.q_proj": 957, "29.k_proj": 381, "29.v_proj": 392, "29.o_proj": 994, "29.gate_proj": 1680, "29.up_proj": 1680, "29.down_proj": 1550, "30.q_proj": 971, "30.k_proj": 389, "30.v_proj": 401, "30.o_proj": 1013, "30.gate_proj": 1680, "30.up_proj": 1680, "30.down_proj": 1593, "31.q_proj": 1053, "31.k_proj": 433, "31.v_proj": 448, "31.o_proj": 1114, "31.gate_proj": 1680, "31.up_proj": 1680, "31.down_proj": 1816 }, "rel_err": { "0.q_proj": 0.10284276306629181, "0.k_proj": 0.09011625498533249, "0.v_proj": 0.4146844744682312, "0.o_proj": 0.4441743791103363, "0.gate_proj": 0.4323641359806061, "0.up_proj": 0.4812891185283661, "0.down_proj": 0.6020375490188599, "1.q_proj": 0.1907663345336914, "1.k_proj": 0.13933585584163666, "1.v_proj": 0.5992577075958252, "1.o_proj": 0.5850644111633301, "1.gate_proj": 0.5018978714942932, "1.up_proj": 0.5401790142059326, "1.down_proj": 0.5769923329353333, "2.q_proj": 0.5321782827377319, "2.k_proj": 0.5425702929496765, "2.v_proj": 0.682027280330658, "2.o_proj": 0.6367513537406921, "2.gate_proj": 0.5818910598754883, "2.up_proj": 0.6077799201011658, "2.down_proj": 0.681037187576294, "3.q_proj": 0.48435133695602417, "3.k_proj": 0.4475176930427551, "3.v_proj": 0.7068923115730286, "3.o_proj": 0.677703857421875, "3.gate_proj": 0.5896062254905701, "3.up_proj": 0.6225286722183228, "3.down_proj": 0.6739526987075806, "4.q_proj": 0.4879390597343445, "4.k_proj": 0.4703967571258545, "4.v_proj": 0.6710249185562134, "4.o_proj": 0.6244789361953735, "4.gate_proj": 0.5910874605178833, "4.up_proj": 0.6321871876716614, "4.down_proj": 0.6802555918693542, "5.q_proj": 0.5210317969322205, "5.k_proj": 0.5119411945343018, "5.v_proj": 0.6989429593086243, "5.o_proj": 0.6632655262947083, "5.gate_proj": 0.5846965312957764, "5.up_proj": 0.6342030763626099, "5.down_proj": 0.6687344908714294, "6.q_proj": 0.505556583404541, "6.k_proj": 0.48692795634269714, "6.v_proj": 0.6872745752334595, "6.o_proj": 0.646428108215332, "6.gate_proj": 0.5995981693267822, "6.up_proj": 0.6370501518249512, "6.down_proj": 0.6577861309051514, "7.q_proj": 0.4832315444946289, "7.k_proj": 0.4549109935760498, "7.v_proj": 0.6812716722488403, "7.o_proj": 0.6252717971801758, "7.gate_proj": 0.5871505737304688, "7.up_proj": 0.6294843554496765, "7.down_proj": 0.654327392578125, "8.q_proj": 0.49038824439048767, "8.k_proj": 0.4526974856853485, "8.v_proj": 0.6452953219413757, "8.o_proj": 0.6178920269012451, "8.gate_proj": 0.5910043716430664, "8.up_proj": 0.6263056993484497, "8.down_proj": 0.647315502166748, "9.q_proj": 0.48501771688461304, "9.k_proj": 0.4705408811569214, "9.v_proj": 0.6292407512664795, "9.o_proj": 0.6174561977386475, "9.gate_proj": 0.5865891575813293, "9.up_proj": 0.6233162879943848, "9.down_proj": 0.6318148970603943, "10.q_proj": 0.4836317300796509, "10.k_proj": 0.4697504937648773, "10.v_proj": 0.6263429522514343, "10.o_proj": 0.6223848462104797, "10.gate_proj": 0.5854211449623108, "10.up_proj": 0.6197803020477295, "10.down_proj": 0.6372787356376648, "11.q_proj": 0.4799323081970215, "11.k_proj": 0.4386899471282959, "11.v_proj": 0.6085149049758911, "11.o_proj": 0.6194866895675659, "11.gate_proj": 0.585040271282196, "11.up_proj": 0.6186556816101074, "11.down_proj": 0.6358811259269714, "12.q_proj": 0.46628451347351074, "12.k_proj": 0.4467337131500244, "12.v_proj": 0.6398075222969055, "12.o_proj": 0.602120578289032, "12.gate_proj": 0.5809478759765625, "12.up_proj": 0.6136550307273865, "12.down_proj": 0.6333811283111572, "13.q_proj": 0.49495774507522583, "13.k_proj": 0.4739055931568146, "13.v_proj": 0.6350283026695251, "13.o_proj": 0.6104384660720825, "13.gate_proj": 0.5884830951690674, "13.up_proj": 0.6139264702796936, "13.down_proj": 0.6283634305000305, "14.q_proj": 0.4639151692390442, "14.k_proj": 0.44637247920036316, "14.v_proj": 0.6475556492805481, "14.o_proj": 0.6104862093925476, "14.gate_proj": 0.5859494209289551, "14.up_proj": 0.6189398169517517, "14.down_proj": 0.6564309597015381, "15.q_proj": 0.46387919783592224, "15.k_proj": 0.42992931604385376, "15.v_proj": 0.675900399684906, "15.o_proj": 0.6175917387008667, "15.gate_proj": 0.5837653279304504, "15.up_proj": 0.6173660159111023, "15.down_proj": 0.6548573970794678, "16.q_proj": 0.4708837568759918, "16.k_proj": 0.4432294964790344, "16.v_proj": 0.6696433424949646, "16.o_proj": 0.6041324138641357, "16.gate_proj": 0.5891684293746948, "16.up_proj": 0.626923680305481, "16.down_proj": 0.656101644039154, "17.q_proj": 0.5025116801261902, "17.k_proj": 0.48880496621131897, "17.v_proj": 0.6514131426811218, "17.o_proj": 0.6207662224769592, "17.gate_proj": 0.6006857752799988, "17.up_proj": 0.6366637349128723, "17.down_proj": 0.6493645906448364, "18.q_proj": 0.4747740626335144, "18.k_proj": 0.4571532607078552, "18.v_proj": 0.6963456869125366, "18.o_proj": 0.5963596701622009, "18.gate_proj": 0.6004551649093628, "18.up_proj": 0.6350386142730713, "18.down_proj": 0.6672616004943848, "19.q_proj": 0.5268400311470032, "19.k_proj": 0.47958895564079285, "19.v_proj": 0.7042666077613831, "19.o_proj": 0.6199771165847778, "19.gate_proj": 0.6060331463813782, "19.up_proj": 0.6367842555046082, "19.down_proj": 0.6811782121658325, "20.q_proj": 0.5465222001075745, "20.k_proj": 0.5119665265083313, "20.v_proj": 0.6899744272232056, "20.o_proj": 0.6347531080245972, "20.gate_proj": 0.6125325560569763, "20.up_proj": 0.636976957321167, "20.down_proj": 0.6877227425575256, "21.q_proj": 0.5809774398803711, "21.k_proj": 0.5469288229942322, "21.v_proj": 0.6610574722290039, "21.o_proj": 0.6689409017562866, "21.gate_proj": 0.6139337420463562, "21.up_proj": 0.6391165256500244, "21.down_proj": 0.6832785606384277, "22.q_proj": 0.6217024922370911, "22.k_proj": 0.5693233609199524, "22.v_proj": 0.6632723808288574, "22.o_proj": 0.6856058835983276, "22.gate_proj": 0.6192406415939331, "22.up_proj": 0.6405629515647888, "22.down_proj": 0.6809287667274475, "23.q_proj": 0.6093561053276062, "23.k_proj": 0.5737671256065369, "23.v_proj": 0.6871448159217834, "23.o_proj": 0.6642407178878784, "23.gate_proj": 0.6219133138656616, "23.up_proj": 0.6424050331115723, "23.down_proj": 0.6742445826530457, "24.q_proj": 0.6159307360649109, "24.k_proj": 0.5868129134178162, "24.v_proj": 0.6966824531555176, "24.o_proj": 0.6866542100906372, "24.gate_proj": 0.6238306760787964, "24.up_proj": 0.6446942687034607, "24.down_proj": 0.6646919250488281, "25.q_proj": 0.6210862994194031, "25.k_proj": 0.5705580115318298, "25.v_proj": 0.679048478603363, "25.o_proj": 0.6676820516586304, "25.gate_proj": 0.6258559823036194, "25.up_proj": 0.6465917825698853, "25.down_proj": 0.6531879901885986, "26.q_proj": 0.6165540218353271, "26.k_proj": 0.5607565641403198, "26.v_proj": 0.6618245840072632, "26.o_proj": 0.6507518887519836, "26.gate_proj": 0.6270332336425781, "26.up_proj": 0.6463326811790466, "26.down_proj": 0.6390087008476257, "27.q_proj": 0.6284988522529602, "27.k_proj": 0.6069222092628479, "27.v_proj": 0.6805939674377441, "27.o_proj": 0.6412052512168884, "27.gate_proj": 0.6305440664291382, "27.up_proj": 0.6479936242103577, "27.down_proj": 0.6380788683891296, "28.q_proj": 0.5876466035842896, "28.k_proj": 0.5294674634933472, "28.v_proj": 0.6338029503822327, "28.o_proj": 0.6242429614067078, "28.gate_proj": 0.6301671266555786, "28.up_proj": 0.6454077959060669, "28.down_proj": 0.6476228833198547, "29.q_proj": 0.5711690783500671, "29.k_proj": 0.5322142243385315, "29.v_proj": 0.670903205871582, "29.o_proj": 0.6061508655548096, "29.gate_proj": 0.6246752142906189, "29.up_proj": 0.6391395926475525, "29.down_proj": 0.6555037498474121, "30.q_proj": 0.5478631854057312, "30.k_proj": 0.4874037206172943, "30.v_proj": 0.6046014428138733, "30.o_proj": 0.5960355997085571, "30.gate_proj": 0.6242106556892395, "30.up_proj": 0.6300148367881775, "30.down_proj": 0.6422404646873474, "31.q_proj": 0.49975186586380005, "31.k_proj": 0.44580137729644775, "31.v_proj": 0.5610269904136658, "31.o_proj": 0.5210819840431213, "31.gate_proj": 0.6289830803871155, "31.up_proj": 0.6037687659263611, "31.down_proj": 0.6013435125350952 }, "step": "swift_svd_dynamic_finetuned", "compression": { "upstream": "hiahei/Swift-SVD", "step": "swift_svd_dynamic", "alloc": "dynamic", "alpha": 0.6, "nsamples": 256, "seqlen": 2048, "dataset": "WikiText2", "seed": 42, "gqa_budget": "shape", "param_fraction": 0.5002723106971154, "mean_rel_err": 0.5875181926946554, "max_rel_err": 0.7061887383460999 }, "factorisation": { "max_rel_err": 0.0014601770089939237, "factor_params": 3491561472, "dense_params": 6979321856 }, "finetune": { "protocol": "svdllm_stage2_u_then_v", "script": "gitclones/SVD-LLM/utils/LoRA.py", "data": "alpaca-cleaned", "epochs_per_half": 2, "learning_rate": 0.0001, "batch_size": 64, "micro_batch_size": 4, "lora_r": 8, "lora_alpha": 16, "lora_dropout": 0.05, "cutoff_len": 256, "val_set_size": 2000, "max_train_samples": "all", "precision": "bf16 (upstream hard-codes fp16; see run_with_torch_compat)" } }