-
Add RewardBench 2 result: 25.3 average (random floor) with full protocol - the OOD boundary measured on a third-party benchmark
-
Fix tokenizer_config compatibility: drop list-form extra_special_tokens (crashes transformers 4.x; tokens remain declared in their standard fields)
-
Add reward_train_max_length: 512 so the calibration regime ships with the weights
-
Fix usage snippet: pairwise comparison, not absolute score; add chance floor + public baseline rows
-
Upload README.md with huggingface_hub
-
Upload tokenizer
-
Upload DebertaV2ForSequenceClassification
-
initial commit
