metadata.json

{
  "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"
  }
}