ArtemVazhentsev21/uhead_math500_gpt_oss_20b_epoch2_best_f1

GPT-OSS-20B MATH-500 uncertainty head (epoch 2)

Uncertainty head trained on claim-level labels for openai/gpt-oss-20b. This artifact is the head saved after epoch 2 while retaining the learning-rate schedule configured for the original 20-epoch run.

Validation metrics on the run split (50 examples, 190 claim labels):

• ROC-AUC: 0.803329
• PR-AUC: 0.342278
• F1: 0.433498
• Accuracy: 0.697368
• Precision: 0.293333
• Recall: 0.830189

Training used seed 42, maximum input length 8192, claim cap 20, positive-class weight 3, and the final hidden layer. Metrics are from the internal validation split, not an independent held-out benchmark.