| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263 |
- {
- "timestamp": "2026-03-19T19:32:47.539221",
- "config": {
- "models": [
- "llama3.2-3b-instruct",
- "Mistral-7B-v0.3",
- "llama3-8b",
- "Qwen2.5-7B-Instruct",
- "gemma-2-9b-it"
- ],
- "model_paths": {
- "llama3.2-3b-instruct": "model/weights/llama3.2-3b-instruct",
- "Mistral-7B-v0.3": "model/weights/Mistral-7B-v0.3",
- "llama3-8b": "model/weights/llama3-8b",
- "Qwen2.5-7B-Instruct": "model/weights/Qwen2.5-7B-Instruct",
- "gemma-2-9b-it": "model/weights/gemma-2-9b-it"
- },
- "datasets": [
- "gsm8k",
- "math",
- "humaneval",
- "alpaca",
- "flores200"
- ],
- "output_dir": "experiments/output/cross_model",
- "batch_size": 4,
- "max_length": 512,
- "pooling": "mean",
- "layer": -1,
- "device": "cuda"
- },
- "r_effs": {
- "gsm8k": {
- "llama3.2-3b-instruct": 4.582842826843262,
- "Mistral-7B-v0.3": 6.044153690338135,
- "llama3-8b": 5.306855201721191,
- "Qwen2.5-7B-Instruct": 8.82113265991211,
- "gemma-2-9b-it": 4.314672470092773
- },
- "math": {
- "llama3.2-3b-instruct": 4.235454559326172,
- "Mistral-7B-v0.3": 5.4554123878479,
- "llama3-8b": 4.868597507476807,
- "Qwen2.5-7B-Instruct": 12.218155860900879,
- "gemma-2-9b-it": 3.9840540885925293
- },
- "humaneval": {
- "llama3.2-3b-instruct": 3.2891879081726074,
- "Mistral-7B-v0.3": 3.4817843437194824,
- "llama3-8b": 3.509075164794922,
- "Qwen2.5-7B-Instruct": 2.463521718978882,
- "gemma-2-9b-it": 2.9243171215057373
- },
- "alpaca": {
- "llama3.2-3b-instruct": 7.3951592445373535,
- "Mistral-7B-v0.3": 16.794479370117188,
- "llama3-8b": 15.56032657623291,
- "Qwen2.5-7B-Instruct": 12.141366004943848,
- "gemma-2-9b-it": 11.04970645904541
- },
- "flores200": {
- "llama3.2-3b-instruct": 7.007132053375244,
- "Mistral-7B-v0.3": 6.034575939178467,
- "llama3-8b": 6.646492958068848,
- "Qwen2.5-7B-Instruct": 4.030577182769775,
- "gemma-2-9b-it": 4.7486114501953125
- }
- },
- "rs_cross": {
- "gsm8k": 2.6233488810620114,
- "math": 9.46074770563815,
- "humaneval": 0.15584515141964855,
- "alpaca": 11.212123561386544,
- "flores200": 1.2816568422497403
- },
- "spd_matrices": {
- "gsm8k": [
- [
- 0.0,
- 0.9500933888470225,
- 0.6383731096474029,
- 4.611884240673232,
- 0.8891039575005614
- ],
- [
- 0.9500933888470225,
- 0.0,
- 0.31697099460200095,
- 3.6617908518262112,
- 1.839189932420268
- ],
- [
- 0.6383731096474029,
- 0.31697099460200095,
- 0.0,
- 3.973511131025831,
- 1.5274690515850768
- ],
- [
- 4.611884240673232,
- 3.6617908518262112,
- 3.973511131025831,
- 0.0,
- 5.500980182610908
- ],
- [
- 0.8891039575005614,
- 1.839189932420268,
- 1.5274690515850768,
- 5.500980182610908,
- 0.0
- ]
- ],
- "math": [
- [
- 0.0,
- 0.6191046293013139,
- 0.38072129363264196,
- 8.854198236824583,
- 0.7776612481108202
- ],
- [
- 0.6191046293013139,
- 0.0,
- 0.23997678527427269,
- 8.236138572882496,
- 1.3957209120529073
- ],
- [
- 0.38072129363264196,
- 0.23997678527427269,
- 0.0,
- 8.473562991325572,
- 1.1582964936098308
- ],
- [
- 8.854198236824583,
- 8.236138572882496,
- 8.473562991325572,
- 0.0,
- 9.631859484935402
- ],
- [
- 0.7776612481108202,
- 1.3957209120529073,
- 1.1582964936098308,
- 9.631859484935402,
- 0.0
- ]
- ],
- "humaneval": [
- [
- 0.0,
- 0.22768383649865587,
- 0.1646244999984714,
- 1.9560668415868125,
- 0.4323351338036199
- ],
- [
- 0.22768383649865587,
- 0.0,
- 0.11305487139532466,
- 1.796532316498885,
- 0.29021389552661636
- ],
- [
- 0.1646244999984714,
- 0.11305487139532466,
- 0.0,
- 1.9078119285905526,
- 0.39178022204703766
- ],
- [
- 1.9560668415868125,
- 1.796532316498885,
- 1.9078119285905526,
- 0.0,
- 1.5238081998406274
- ],
- [
- 0.4323351338036199,
- 0.29021389552661636,
- 0.39178022204703766,
- 1.5238081998406274,
- 0.0
- ]
- ],
- "alpaca": [
- [
- 0.0,
- 3.245752073437664,
- 5.19263677108836,
- 1.9810623051768308,
- 4.3227948398617
- ],
- [
- 3.245752073437664,
- 0.0,
- 2.200825753755322,
- 3.7746811947154533,
- 2.2446554958004676
- ],
- [
- 5.19263677108836,
- 2.200825753755322,
- 0.0,
- 5.721565892366148,
- 1.1033485618916488
- ],
- [
- 1.9810623051768308,
- 3.7746811947154533,
- 5.721565892366148,
- 0.0,
- 5.184835513324546
- ],
- [
- 4.3227948398617,
- 2.2446554958004676,
- 1.1033485618916488,
- 5.184835513324546,
- 0.0
- ]
- ],
- "flores200": [
- [
- 0.0,
- 0.7929460690446842,
- 0.21997497437804503,
- 2.6852593631365975,
- 1.866262298807369
- ],
- [
- 0.7929460690446842,
- 0.0,
- 0.5730167413137194,
- 1.892317294979423,
- 1.0733175818346883
- ],
- [
- 0.21997497437804503,
- 0.5730167413137194,
- 0.0,
- 2.4652978710671136,
- 1.6462923097836648
- ],
- [
- 2.6852593631365975,
- 1.892317294979423,
- 2.4652978710671136,
- 0.0,
- 0.8191072943568687
- ],
- [
- 1.866262298807369,
- 1.0733175818346883,
- 1.6462923097836648,
- 0.8191072943568687,
- 0.0
- ]
- ]
- }
- }
|