| { |
| "format_version": 1, |
| "artifacts": { |
| "mlpackage": "GLiNERRelexBase.mlpackage", |
| "mlmodelc": "GLiNERRelexBase.mlmodelc" |
| }, |
| "tokenizer": { |
| "type": "DebertaV2 (SentencePiece unigram), is_split_into_words semantics", |
| "files": { |
| "tokenizer_json": "tokenizer.json", |
| "sentencepiece_model": "spm.model", |
| "tokenizer_config": "tokenizer_config.json", |
| "added_tokens": "added_tokens.json", |
| "special_tokens_map": "special_tokens_map.json" |
| }, |
| "special_token_ids": { |
| "pad": 0, |
| "cls": 1, |
| "sep": 2, |
| "unk": 3, |
| "mask": 128000, |
| "ent_prompt": 128001, |
| "sep_prompt": 128002, |
| "rel_prompt": 128003 |
| } |
| }, |
| "prompt_construction": { |
| "description": "Word-level prompt prepended to the whitespace-split text words, then tokenized as pre-split words ([CLS] + subtokens + [SEP], each word SentencePiece-encoded with a leading \u2581).", |
| "word_layout": "[<<ENT>>, label_1, <<ENT>>, label_2, ..., <<SEP>>, <<REL>>, relation_1, <<REL>>, relation_2, ..., <<SEP>>, text_word_1, ..., text_word_L]", |
| "words_splitter_regex": "\\w+(?:[-_]\\w+)*|\\S", |
| "notes": [ |
| "Both <<SEP>> tokens are required: one closes the entity-label block, one closes the relation block.", |
| "Labels/relations must be deduplicated preserving order (dict.fromkeys semantics).", |
| "The k-th <<ENT>> token feeds entity-class slot k-1; the k-th <<REL>> token feeds relation-class slot k-1." |
| ] |
| }, |
| "inputs": { |
| "input_ids": { |
| "shape": [ |
| 1, |
| 512 |
| ], |
| "dtype": "int32", |
| "semantics": "DeBERTa subtoken ids, [PAD]=0 right-padded" |
| }, |
| "attention_mask": { |
| "shape": [ |
| 1, |
| 512 |
| ], |
| "dtype": "int32", |
| "semantics": "1 for real subtokens (incl. CLS/SEP), 0 for padding" |
| }, |
| "words_mask": { |
| "shape": [ |
| 1, |
| 512 |
| ], |
| "dtype": "int32", |
| "semantics": "1-based text-word index on the FIRST subtoken of each text word; 0 on prompt words, special tokens, continuation subtokens and padding" |
| }, |
| "text_lengths": { |
| "shape": [ |
| 1, |
| 1 |
| ], |
| "dtype": "int32", |
| "semantics": "number of text words L (excludes prompt words)" |
| }, |
| "span_idx": { |
| "shape": [ |
| 1, |
| 16, |
| 2 |
| ], |
| "dtype": "int32", |
| "semantics": "word-level [start, end] (inclusive) of extracted candidate entities for the relation pass; zeros elsewhere" |
| }, |
| "span_mask": { |
| "shape": [ |
| 1, |
| 16 |
| ], |
| "dtype": "int32", |
| "semantics": "1 for occupied entity slots, else 0" |
| } |
| }, |
| "outputs": { |
| "logits": { |
| "shape": [ |
| 1, |
| 320, |
| 12, |
| 3 |
| ], |
| "dtype": "float32", |
| "axes": [ |
| "batch", |
| "word position (only w < text_lengths valid)", |
| "entity class slot (only c < num_labels valid)", |
| "BIO component: 0=start, 1=end, 2=inside" |
| ], |
| "semantics": "raw token BIO logits; apply sigmoid before thresholding" |
| }, |
| "rel_logits": { |
| "shape": [ |
| 1, |
| 240, |
| 8 |
| ], |
| "dtype": "float32", |
| "axes": [ |
| "batch", |
| "ordered entity-slot pair", |
| "relation class slot (only r < num_relations valid)" |
| ], |
| "pair_slot_formula": "slot(i, j) = i * (MAX_ENT_SLOTS - 1) + (j if j < i else j - 1), i != j", |
| "semantics": "raw relation logits for pair (head=slot i, tail=slot j); apply sigmoid; only pairs with both slots valid in span_mask are meaningful" |
| } |
| }, |
| "static_shapes": { |
| "seq_len": 512, |
| "max_words": 320, |
| "max_entity_types": 12, |
| "max_relation_types": 8, |
| "max_entity_slots": 16, |
| "num_pairs": 240, |
| "hidden_size": 768 |
| }, |
| "decode_algorithm": { |
| "protocol": "two-pass: pass1 span_idx/span_mask all zero -> logits; extract candidate spans; pass2 with spans -> rel_logits", |
| "entity_threshold": 0.4, |
| "relation_threshold": 0.7, |
| "flat_ner": false, |
| "multi_label": false, |
| "candidate_span_extraction": "For every (word, class) with sigmoid(start) > t and every (word', class) with sigmoid(end) > t, word <= word', keep (word, word') if sigmoid(inside) > t for the whole run. Iterate starts row-major (word-major) then ends; keep duplicates from different classes. These ordered spans fill span_idx slots 0..E-1.", |
| "entity_decoding": "Same start/end/inside pairing on sigmoid probabilities; span score = min(start, end, all inside); then greedy overlap filtering: sort by score desc (stable), keep spans that do not conflict (flat_ner=false allows nesting, identical boundaries conflict unless multi_label); finally sort by start (stable).", |
| "relation_decoding": "Map each candidate slot to the FIRST decoded entity with identical word boundaries (slots without a match are skipped). For each ordered pair (i, j), i != j, over slots 0..E-1 in row-major order, emit (head, relation_r, tail, sigmoid) for every relation class r with sigmoid(rel_logits[slot(i,j), r]) > relation_threshold. Duplicate-boundary slots intentionally emit duplicate relations (PyTorch decoder behavior)." |
| }, |
| "limits": { |
| "max_text_words": 320, |
| "max_subtokens": 512, |
| "max_labels": 12, |
| "max_relations": 8, |
| "max_candidate_spans": 16, |
| "unused_class_columns": "logit columns for class slots beyond the provided label count contain garbage (zero prompt embeddings) and MUST be ignored" |
| }, |
| "sha256": { |
| "mlpackage_tree": "9d948eb9e9c690ee5941ab28261361dad00ce74d70f997ca35ac9e09a6e4ab0a", |
| "mlmodelc_tree": "f700158783796a7b93e366708c8467899ac6d4c7e18da567537046e1dfa43a5f" |
| } |
| } |
