| { | |
| "arch": "CorefHead (Lee-2017 antecedent scorer on the SHARED frozen mDeBERTa backbone)", | |
| "precision": "fp16", | |
| "hidden": 768, | |
| "proj": 256, | |
| "mention_rep": "[start_tok ; end_tok ; mean(span)] -> proj(768*3->256)", | |
| "pair_score": "MLP([m_i ; m_j ; m_i*m_j]) + learned dummy antecedent", | |
| "decode": "argmax antecedent per mention -> union-find clusters (dummy=no antecedent)", | |
| "compose": "rewrite-first at inference (resolve clusters, swap anaphora to cluster head, offset back-map keeps provenance)", | |
| "trained_on": "research/data/coref (4400 train / 1100 val, EN/JA/KO/TH/ZH)", | |
| "val_link_accuracy": "97.8%", | |
| "weights": "coref_head_fp16.pt (mention encoder + pair MLP + dummy)", | |
| "backbone": "shares model_fp16.safetensors encoder (frozen)" | |
| } |