e1_result.json 5.7 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263
  1. {
  2. "timestamp": "2026-03-19T19:32:47.539221",
  3. "config": {
  4. "models": [
  5. "llama3.2-3b-instruct",
  6. "Mistral-7B-v0.3",
  7. "llama3-8b",
  8. "Qwen2.5-7B-Instruct",
  9. "gemma-2-9b-it"
  10. ],
  11. "model_paths": {
  12. "llama3.2-3b-instruct": "model/weights/llama3.2-3b-instruct",
  13. "Mistral-7B-v0.3": "model/weights/Mistral-7B-v0.3",
  14. "llama3-8b": "model/weights/llama3-8b",
  15. "Qwen2.5-7B-Instruct": "model/weights/Qwen2.5-7B-Instruct",
  16. "gemma-2-9b-it": "model/weights/gemma-2-9b-it"
  17. },
  18. "datasets": [
  19. "gsm8k",
  20. "math",
  21. "humaneval",
  22. "alpaca",
  23. "flores200"
  24. ],
  25. "output_dir": "experiments/output/cross_model",
  26. "batch_size": 4,
  27. "max_length": 512,
  28. "pooling": "mean",
  29. "layer": -1,
  30. "device": "cuda"
  31. },
  32. "r_effs": {
  33. "gsm8k": {
  34. "llama3.2-3b-instruct": 4.582842826843262,
  35. "Mistral-7B-v0.3": 6.044153690338135,
  36. "llama3-8b": 5.306855201721191,
  37. "Qwen2.5-7B-Instruct": 8.82113265991211,
  38. "gemma-2-9b-it": 4.314672470092773
  39. },
  40. "math": {
  41. "llama3.2-3b-instruct": 4.235454559326172,
  42. "Mistral-7B-v0.3": 5.4554123878479,
  43. "llama3-8b": 4.868597507476807,
  44. "Qwen2.5-7B-Instruct": 12.218155860900879,
  45. "gemma-2-9b-it": 3.9840540885925293
  46. },
  47. "humaneval": {
  48. "llama3.2-3b-instruct": 3.2891879081726074,
  49. "Mistral-7B-v0.3": 3.4817843437194824,
  50. "llama3-8b": 3.509075164794922,
  51. "Qwen2.5-7B-Instruct": 2.463521718978882,
  52. "gemma-2-9b-it": 2.9243171215057373
  53. },
  54. "alpaca": {
  55. "llama3.2-3b-instruct": 7.3951592445373535,
  56. "Mistral-7B-v0.3": 16.794479370117188,
  57. "llama3-8b": 15.56032657623291,
  58. "Qwen2.5-7B-Instruct": 12.141366004943848,
  59. "gemma-2-9b-it": 11.04970645904541
  60. },
  61. "flores200": {
  62. "llama3.2-3b-instruct": 7.007132053375244,
  63. "Mistral-7B-v0.3": 6.034575939178467,
  64. "llama3-8b": 6.646492958068848,
  65. "Qwen2.5-7B-Instruct": 4.030577182769775,
  66. "gemma-2-9b-it": 4.7486114501953125
  67. }
  68. },
  69. "rs_cross": {
  70. "gsm8k": 2.6233488810620114,
  71. "math": 9.46074770563815,
  72. "humaneval": 0.15584515141964855,
  73. "alpaca": 11.212123561386544,
  74. "flores200": 1.2816568422497403
  75. },
  76. "spd_matrices": {
  77. "gsm8k": [
  78. [
  79. 0.0,
  80. 0.9500933888470225,
  81. 0.6383731096474029,
  82. 4.611884240673232,
  83. 0.8891039575005614
  84. ],
  85. [
  86. 0.9500933888470225,
  87. 0.0,
  88. 0.31697099460200095,
  89. 3.6617908518262112,
  90. 1.839189932420268
  91. ],
  92. [
  93. 0.6383731096474029,
  94. 0.31697099460200095,
  95. 0.0,
  96. 3.973511131025831,
  97. 1.5274690515850768
  98. ],
  99. [
  100. 4.611884240673232,
  101. 3.6617908518262112,
  102. 3.973511131025831,
  103. 0.0,
  104. 5.500980182610908
  105. ],
  106. [
  107. 0.8891039575005614,
  108. 1.839189932420268,
  109. 1.5274690515850768,
  110. 5.500980182610908,
  111. 0.0
  112. ]
  113. ],
  114. "math": [
  115. [
  116. 0.0,
  117. 0.6191046293013139,
  118. 0.38072129363264196,
  119. 8.854198236824583,
  120. 0.7776612481108202
  121. ],
  122. [
  123. 0.6191046293013139,
  124. 0.0,
  125. 0.23997678527427269,
  126. 8.236138572882496,
  127. 1.3957209120529073
  128. ],
  129. [
  130. 0.38072129363264196,
  131. 0.23997678527427269,
  132. 0.0,
  133. 8.473562991325572,
  134. 1.1582964936098308
  135. ],
  136. [
  137. 8.854198236824583,
  138. 8.236138572882496,
  139. 8.473562991325572,
  140. 0.0,
  141. 9.631859484935402
  142. ],
  143. [
  144. 0.7776612481108202,
  145. 1.3957209120529073,
  146. 1.1582964936098308,
  147. 9.631859484935402,
  148. 0.0
  149. ]
  150. ],
  151. "humaneval": [
  152. [
  153. 0.0,
  154. 0.22768383649865587,
  155. 0.1646244999984714,
  156. 1.9560668415868125,
  157. 0.4323351338036199
  158. ],
  159. [
  160. 0.22768383649865587,
  161. 0.0,
  162. 0.11305487139532466,
  163. 1.796532316498885,
  164. 0.29021389552661636
  165. ],
  166. [
  167. 0.1646244999984714,
  168. 0.11305487139532466,
  169. 0.0,
  170. 1.9078119285905526,
  171. 0.39178022204703766
  172. ],
  173. [
  174. 1.9560668415868125,
  175. 1.796532316498885,
  176. 1.9078119285905526,
  177. 0.0,
  178. 1.5238081998406274
  179. ],
  180. [
  181. 0.4323351338036199,
  182. 0.29021389552661636,
  183. 0.39178022204703766,
  184. 1.5238081998406274,
  185. 0.0
  186. ]
  187. ],
  188. "alpaca": [
  189. [
  190. 0.0,
  191. 3.245752073437664,
  192. 5.19263677108836,
  193. 1.9810623051768308,
  194. 4.3227948398617
  195. ],
  196. [
  197. 3.245752073437664,
  198. 0.0,
  199. 2.200825753755322,
  200. 3.7746811947154533,
  201. 2.2446554958004676
  202. ],
  203. [
  204. 5.19263677108836,
  205. 2.200825753755322,
  206. 0.0,
  207. 5.721565892366148,
  208. 1.1033485618916488
  209. ],
  210. [
  211. 1.9810623051768308,
  212. 3.7746811947154533,
  213. 5.721565892366148,
  214. 0.0,
  215. 5.184835513324546
  216. ],
  217. [
  218. 4.3227948398617,
  219. 2.2446554958004676,
  220. 1.1033485618916488,
  221. 5.184835513324546,
  222. 0.0
  223. ]
  224. ],
  225. "flores200": [
  226. [
  227. 0.0,
  228. 0.7929460690446842,
  229. 0.21997497437804503,
  230. 2.6852593631365975,
  231. 1.866262298807369
  232. ],
  233. [
  234. 0.7929460690446842,
  235. 0.0,
  236. 0.5730167413137194,
  237. 1.892317294979423,
  238. 1.0733175818346883
  239. ],
  240. [
  241. 0.21997497437804503,
  242. 0.5730167413137194,
  243. 0.0,
  244. 2.4652978710671136,
  245. 1.6462923097836648
  246. ],
  247. [
  248. 2.6852593631365975,
  249. 1.892317294979423,
  250. 2.4652978710671136,
  251. 0.0,
  252. 0.8191072943568687
  253. ],
  254. [
  255. 1.866262298807369,
  256. 1.0733175818346883,
  257. 1.6462923097836648,
  258. 0.8191072943568687,
  259. 0.0
  260. ]
  261. ]
  262. }
  263. }