Upload folder using huggingface_hub

README.md

@@ -24,7 +24,9 @@ AutoScientist Challenge (Part 2).
 
 ## Measured improvement
 
-AutoScientist reported **best_win_rate = 0.5492** over 2 iterations against `meta-llama/Llama-3.2-3B-Instruct` - i.e. the fine-tuned model is preferred over its own baseline in 54.9% of comparisons. An earlier 5-iteration run of the same dataset reached 0.5674 before the job wedged at iteration 4/5 and had to be cancelled, so 0.5492 is the reproducible, completed figure.
+AutoScientist reported **best_win_rate = 0.6262** against `meta-llama/Llama-3.2-3B-Instruct` - the fine-tuned model is preferred over its own baseline in 62.6% of comparisons.
+
+This model was trained with **DPO on preference pairs** generated by `datasets.run(training_type='preference_pairs')`. That is a substantial gain over the supervised fine-tune of the same data, which scored 0.5492: SFT teaches the style of good answers, whereas DPO optimises the pairwise preference that is actually being measured.
 
 Evaluation methodology, including the position-swap and dual-judge controls, is in
 `EVAL.md` in the project repository. The held-out split used is published alongside the

adapter_config.json

@@ -24,11 +24,11 @@
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
+    "o_proj",
+    "up_proj",
     "gate_proj",
     "v_proj",
-    "up_proj",
     "down_proj",
-    "o_proj",
     "k_proj",
     "q_proj"
   ],

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:50cdd02bcab4ed669c2d113478231d85a5caa3e102a5ca7b7da951ff915d930d
+oid sha256:eca27f188528f4e39b1970cbb6e9e8e636b0c7c1e28b943ffffd8a6f463f4a87
 size 97307544

trainer_state.json

@@ -3,202 +3,230 @@
   "best_metric": null,
   "best_model_checkpoint": null,
   "epoch": 1.0,
-  "eval_steps": 4,
-  "global_step": 21,
+  "eval_steps": 5,
+  "global_step": 25,
   "is_hyper_param_search": false,
   "is_local_process_zero": true,
   "is_world_process_zero": true,
   "log_history": [
     {
-      "epoch": 0.047619047619047616,
-      "grad_norm": 0.26583361625671387,
+      "epoch": 0.04,
+      "grad_norm": 0.3135530948638916,
       "learning_rate": 0.0,
-      "loss": 1.838623046875,
+      "loss": 1.8271484375,
       "step": 1
     },
     {
-      "epoch": 0.09523809523809523,
-      "grad_norm": 0.2549391984939575,
-      "learning_rate": 3.3333333333333333e-06,
-      "loss": 1.7900390625,
+      "epoch": 0.08,
+      "grad_norm": 0.30578845739364624,
+      "learning_rate": 5e-05,
+      "loss": 1.76904296875,
       "step": 2
     },
     {
-      "epoch": 0.14285714285714285,
-      "grad_norm": 0.27876609563827515,
-      "learning_rate": 6.666666666666667e-06,
-      "loss": 1.8592529296875,
+      "epoch": 0.12,
+      "grad_norm": 0.32055363059043884,
+      "learning_rate": 4.980750938091074e-05,
+      "loss": 1.8251953125,
       "step": 3
     },
     {
-      "epoch": 0.19047619047619047,
-      "grad_norm": 0.2896035611629486,
-      "learning_rate": 1e-05,
-      "loss": 1.801513671875,
+      "epoch": 0.16,
+      "grad_norm": 0.3271876871585846,
+      "learning_rate": 4.9233331091504034e-05,
+      "loss": 1.8466796875,
       "step": 4
     },
     {
-      "epoch": 0.23809523809523808,
-      "grad_norm": 0.3665730953216553,
-      "learning_rate": 9.931634888554937e-06,
-      "loss": 1.545654296875,
+      "epoch": 0.2,
+      "grad_norm": 0.3510723114013672,
+      "learning_rate": 4.8287289481503954e-05,
+      "loss": 1.89501953125,
       "step": 5
     },
     {
-      "epoch": 0.23809523809523808,
-      "eval_loss": 1.732421875,
-      "eval_runtime": 2.193,
-      "eval_samples_per_second": 1.368,
-      "eval_steps_per_second": 0.456,
+      "epoch": 0.2,
+      "eval_loss": 1.7734375,
+      "eval_runtime": 2.2485,
+      "eval_samples_per_second": 0.889,
+      "eval_steps_per_second": 0.445,
       "step": 5
     },
     {
-      "epoch": 0.2857142857142857,
-      "grad_norm": 0.2946288287639618,
-      "learning_rate": 9.728616793536588e-06,
-      "loss": 1.818359375,
+      "epoch": 0.24,
+      "grad_norm": 0.3387308716773987,
+      "learning_rate": 4.6985571585149876e-05,
+      "loss": 1.76806640625,
       "step": 6
     },
     {
-      "epoch": 0.3333333333333333,
-      "grad_norm": 0.2675096094608307,
-      "learning_rate": 9.397114317029975e-06,
-      "loss": 1.7734375,
+      "epoch": 0.28,
+      "grad_norm": 0.34001627564430237,
+      "learning_rate": 4.535045015655279e-05,
+      "loss": 1.7548828125,
       "step": 7
     },
     {
-      "epoch": 0.38095238095238093,
-      "grad_norm": 0.2651223838329315,
-      "learning_rate": 8.947199994035402e-06,
-      "loss": 1.78955078125,
+      "epoch": 0.32,
+      "grad_norm": 0.3075582683086395,
+      "learning_rate": 4.340990257669732e-05,
+      "loss": 1.6689453125,
       "step": 8
     },
     {
-      "epoch": 0.42857142857142855,
-      "grad_norm": 0.31145110726356506,
-      "learning_rate": 8.392544243589428e-06,
-      "loss": 1.8505859375,
+      "epoch": 0.36,
+      "grad_norm": 0.29662805795669556,
+      "learning_rate": 4.1197132152696215e-05,
+      "loss": 1.7255859375,
       "step": 9
     },
     {
-      "epoch": 0.42857142857142855,
-      "eval_loss": 1.724609375,
-      "eval_runtime": 2.188,
-      "eval_samples_per_second": 1.371,
-      "eval_steps_per_second": 0.457,
-      "step": 9
+      "epoch": 0.4,
+      "grad_norm": 0.28356656432151794,
+      "learning_rate": 3.875e-05,
+      "loss": 1.7080078125,
+      "step": 10
     },
     {
-      "epoch": 0.47619047619047616,
-      "grad_norm": 0.27500542998313904,
-      "learning_rate": 7.75e-06,
-      "loss": 1.796630859375,
+      "epoch": 0.4,
+      "eval_loss": 1.68359375,
+      "eval_runtime": 2.0485,
+      "eval_samples_per_second": 0.976,
+      "eval_steps_per_second": 0.488,
       "step": 10
     },
     {
-      "epoch": 0.5238095238095238,
-      "grad_norm": 0.27637097239494324,
-      "learning_rate": 7.0390906449655104e-06,
-      "loss": 1.82177734375,
+      "epoch": 0.44,
+      "grad_norm": 0.2615479826927185,
+      "learning_rate": 3.611037722821452e-05,
+      "loss": 1.66796875,
       "step": 11
     },
     {
-      "epoch": 0.5714285714285714,
-      "grad_norm": 0.28143683075904846,
-      "learning_rate": 6.281416799501188e-06,
-      "loss": 1.84130859375,
+      "epoch": 0.48,
+      "grad_norm": 0.25153541564941406,
+      "learning_rate": 3.332342851480672e-05,
+      "loss": 1.6669921875,
       "step": 12
     },
     {
-      "epoch": 0.6190476190476191,
-      "grad_norm": 0.29704588651657104,
-      "learning_rate": 5.500000000000001e-06,
-      "loss": 1.807373046875,
+      "epoch": 0.52,
+      "grad_norm": 0.236228808760643,
+      "learning_rate": 3.0436839324951166e-05,
+      "loss": 1.6611328125,
       "step": 13
     },
     {
-      "epoch": 0.6190476190476191,
-      "eval_loss": 1.71875,
-      "eval_runtime": 2.1801,
-      "eval_samples_per_second": 1.376,
-      "eval_steps_per_second": 0.459,
-      "step": 13
-    },
-    {
-      "epoch": 0.6666666666666666,
-      "grad_norm": 0.2861112654209137,
-      "learning_rate": 4.718583200498814e-06,
-      "loss": 1.81103515625,
+      "epoch": 0.56,
+      "grad_norm": 0.24782223999500275,
+      "learning_rate": 2.7500000000000004e-05,
+      "loss": 1.740234375,
       "step": 14
     },
     {
-      "epoch": 0.7142857142857143,
-      "grad_norm": 0.3324793577194214,
-      "learning_rate": 3.960909355034491e-06,
-      "loss": 1.796875,
+      "epoch": 0.6,
+      "grad_norm": 0.2386341094970703,
+      "learning_rate": 2.4563160675048846e-05,
+      "loss": 1.64404296875,
       "step": 15
     },
     {
-      "epoch": 0.7619047619047619,
-      "grad_norm": 0.25911256670951843,
-      "learning_rate": 3.2500000000000015e-06,
-      "loss": 1.736572265625,
-      "step": 16
+      "epoch": 0.6,
+      "eval_loss": 1.6484375,
+      "eval_runtime": 2.051,
+      "eval_samples_per_second": 0.975,
+      "eval_steps_per_second": 0.488,
+      "step": 15
     },
     {
-      "epoch": 0.8095238095238095,
-      "grad_norm": 0.2930397391319275,
-      "learning_rate": 2.607455756410573e-06,
-      "loss": 1.818359375,
-      "step": 17
+      "epoch": 0.64,
+      "grad_norm": 0.22649618983268738,
+      "learning_rate": 2.1676571485193282e-05,
+      "loss": 1.55908203125,
+      "step": 16
     },
     {
-      "epoch": 0.8095238095238095,
-      "eval_loss": 1.7109375,
-      "eval_runtime": 2.19,
-      "eval_samples_per_second": 1.37,
-      "eval_steps_per_second": 0.457,
+      "epoch": 0.68,
+      "grad_norm": 0.25247156620025635,
+      "learning_rate": 1.888962277178548e-05,
+      "loss": 1.6220703125,
       "step": 17
     },
     {
-      "epoch": 0.8571428571428571,
-      "grad_norm": 0.29004520177841187,
-      "learning_rate": 2.0528000059646e-06,
-      "loss": 1.81201171875,
+      "epoch": 0.72,
+      "grad_norm": 0.2598176896572113,
+      "learning_rate": 1.6250000000000005e-05,
+      "loss": 1.662109375,
       "step": 18
     },
     {
-      "epoch": 0.9047619047619048,
-      "grad_norm": 0.2956332862377167,
-      "learning_rate": 1.602885682970026e-06,
-      "loss": 1.8251953125,
+      "epoch": 0.76,
+      "grad_norm": 0.23730042576789856,
+      "learning_rate": 1.3802867847303785e-05,
+      "loss": 1.5947265625,
       "step": 19
     },
     {
-      "epoch": 0.9523809523809523,
-      "grad_norm": 0.3092670440673828,
-      "learning_rate": 1.2713832064634127e-06,
-      "loss": 1.779296875,
+      "epoch": 0.8,
+      "grad_norm": 0.24574804306030273,
+      "learning_rate": 1.1590097423302684e-05,
+      "loss": 1.69921875,
       "step": 20
     },
     {
-      "epoch": 1.0,
-      "grad_norm": 0.29927435517311096,
-      "learning_rate": 1.0683651114450641e-06,
-      "loss": 1.79345703125,
+      "epoch": 0.8,
+      "eval_loss": 1.625,
+      "eval_runtime": 2.0556,
+      "eval_samples_per_second": 0.973,
+      "eval_steps_per_second": 0.486,
+      "step": 20
+    },
+    {
+      "epoch": 0.84,
+      "grad_norm": 0.23554882407188416,
+      "learning_rate": 9.649549843447212e-06,
+      "loss": 1.65576171875,
       "step": 21
     },
+    {
+      "epoch": 0.88,
+      "grad_norm": 0.24011166393756866,
+      "learning_rate": 8.01442841485013e-06,
+      "loss": 1.625,
+      "step": 22
+    },
+    {
+      "epoch": 0.92,
+      "grad_norm": 0.24006858468055725,
+      "learning_rate": 6.712710518496049e-06,
+      "loss": 1.611328125,
+      "step": 23
+    },
+    {
+      "epoch": 0.96,
+      "grad_norm": 0.22434408962726593,
+      "learning_rate": 5.766668908495966e-06,
+      "loss": 1.5771484375,
+      "step": 24
+    },
     {
       "epoch": 1.0,
-      "eval_loss": 1.7109375,
-      "eval_runtime": 2.1924,
-      "eval_samples_per_second": 1.368,
-      "eval_steps_per_second": 0.456,
-      "step": 21
+      "grad_norm": 0.21985848248004913,
+      "learning_rate": 5.192490619089267e-06,
+      "loss": 1.5849609375,
+      "step": 25
+    },
+    {
+      "epoch": 1.0,
+      "eval_loss": 1.62109375,
+      "eval_runtime": 2.0583,
+      "eval_samples_per_second": 0.972,
+      "eval_steps_per_second": 0.486,
+      "step": 25
     }
   ],
   "logging_steps": 1.0,
-  "max_steps": 21,
+  "max_steps": 25,
   "num_input_tokens_seen": 0,
   "num_train_epochs": 1,
   "save_steps": 0,
@@ -214,7 +242,7 @@
       "attributes": {}
     }
   },
-  "total_flos": 1.722081761694843e+17,
+  "total_flos": 2.8168959689752576e+16,
   "train_batch_size": 1,
   "trial_name": null,
   "trial_params": null