v2: alias-list training and Gutenberg self-training (29,214 teacher-labelled windows)
8fd1c33 verified | { | |
| "window": 510, | |
| "inputs": [ | |
| "input_ids", | |
| "attention_mask", | |
| "mention_starts", | |
| "mention_ends", | |
| "quote_starts", | |
| "quote_ends" | |
| ], | |
| "outputs": [ | |
| "entity_logits", | |
| "nominal_logits", | |
| "coref_scores", | |
| "attribution_scores" | |
| ], | |
| "notes": "coref_scores[i][j] scores mention j as antecedent of mention i (j < i); 0 is the no-antecedent score. attribution_scores[q][0] is 'speaker not in window', column m+1 is mention m.", | |
| "training": { | |
| "base": "roberta-base", | |
| "teacherBase": "microsoft/deberta-v3-large", | |
| "sources": [ | |
| "litbank", | |
| "pdnc", | |
| "pdnc_alias", | |
| "gutenberg" | |
| ], | |
| "documents": 272, | |
| "folds": 5, | |
| "epochs": 3, | |
| "lr": 3e-05, | |
| "teacherEpochs": 2, | |
| "teacherLr": 2e-05, | |
| "distillWeight": 1.0, | |
| "lossWeights": { | |
| "entity": 3.0, | |
| "nominal": 2.0 | |
| }, | |
| "novelWindows": 0, | |
| "litbankRepeat": 4, | |
| "window": 510, | |
| "offsets": [ | |
| 0, | |
| 255 | |
| ], | |
| "teacherOffsets": [ | |
| 0 | |
| ], | |
| "unlabeledWindows": 30000, | |
| "unlabeledDocuments": 100, | |
| "seed": 1, | |
| "teacherCheckpoint": null | |
| } | |
| } | |