| { |
| "manifest_version": 1, |
| "role": "reference_segmenter_text", |
| "name": "reference-segmenter-text-mmbert", |
| "version": "sha256-a9a5f0ce3ff8", |
| "task": "token_classification", |
| "runtime": { |
| "engine": "onnxruntime", |
| "requires": [ |
| "onnxruntime", |
| "tokenizers", |
| "numpy" |
| ], |
| "opset": null, |
| "executes_code_on_load": false |
| }, |
| "files": [ |
| { |
| "path": "onnx/model.onnx", |
| "bytes": 268620732, |
| "sha256": "a9a5f0ce3ff8f398ab6517ae67e09f06fb5610015f5a5307b5b86774ad5e8c7a", |
| "kind": "model" |
| }, |
| { |
| "path": "onnx/segmenter.json", |
| "bytes": 500, |
| "sha256": "964470334fc8f5a6d407f4d1961b1349fafca22625a2c7f227776d212b072b9f", |
| "kind": "legacy_manifest" |
| }, |
| { |
| "path": "onnx/tokenizer.json", |
| "bytes": 34363188, |
| "sha256": "609d8f4c067cd3950f88594c5a802616cea245823836ef5848ee4fc40aab5b6f", |
| "kind": "tokenizer" |
| } |
| ], |
| "encoder": { |
| "id": "jhu-clsp/mmBERT-small", |
| "revision": null |
| }, |
| "labels": { |
| "tags": [ |
| "O", |
| "B-REF", |
| "I-REF" |
| ] |
| }, |
| "input": { |
| "layout": "single_text", |
| "max_length": null, |
| "tokenizer_file": "onnx/tokenizer.json", |
| "window": 1024, |
| "stride": 768, |
| "add_special_tokens": false, |
| "truncation": null, |
| "template": null, |
| "normalization": null, |
| "pair": null, |
| "features": null, |
| "special_tokens": {}, |
| "model_inputs": [ |
| { |
| "name": "input_ids", |
| "dtype": "int64", |
| "shape": [ |
| "b", |
| "t" |
| ] |
| }, |
| { |
| "name": "attention_mask", |
| "dtype": "int64", |
| "shape": [ |
| "b", |
| "t" |
| ] |
| } |
| ], |
| "model_outputs": [ |
| { |
| "name": "logits", |
| "dtype": "float32", |
| "shape": [ |
| "b", |
| "t", |
| 3 |
| ] |
| } |
| ] |
| }, |
| "decode": { |
| "method": "window_argmax_merge", |
| "crf": null, |
| "thresholds": {}, |
| "when_nothing_found": "first_unit_forced", |
| "calibration": null |
| }, |
| "quantization": { |
| "weights": "int8_embeddings", |
| "note": "token-embedding table stored as int8 with one fp32 scale per row" |
| }, |
| "provenance": { |
| "ledger_key": "reference-segmenter-text-mmbert", |
| "run_record_sha256": "b5a3cf9c0821f0b674c4e13ba45a9431b984197509c2c4980c6262398ab9fb13", |
| "platform_commit": "0b248015bc8c8f858409d8919f95ad6979dd158a", |
| "registry_build": "2026-11-10T21:38:59Z#d8d427de3ea7", |
| "split_id": "corpus-split-v1@c6fc246571f4" |
| }, |
| "licence": "mit", |
| "compat": { |
| "bibr_min_version": null, |
| "export_schema": null |
| }, |
| "self_test": { |
| "tolerance": { |
| "probability_abs": 0.01, |
| "logit_abs": 0.05, |
| "reason": "onnxruntime builds, CPUs and BLAS kernels give slightly different float results; the largest difference measured between onnxruntime 1.27.0 and 1.31.0 on the eight bundles normalised so far was 8e-05 (stored values are rounded to 5 places), so these limits are about two orders of magnitude looser than the drift seen, and far tighter than any change of a decision; tags and label lists are compared exactly" |
| }, |
| "cases": [ |
| { |
| "name": "three numbered references", |
| "input": { |
| "text": "References\n[1] Marlowe J. Invented estuaries. J Fict Hydrol 2018;14:101-117.\n[2] Okonkwo T. Notes on lighthouses. Imaginary Press; 1977.\n[3] Brandt E. A ledger of imaginary tides. Fict Rev 1999;3:5-9." |
| }, |
| "expect": { |
| "tokens": 86, |
| "tags": [ |
| "I-REF", |
| "I-REF", |
| "B-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "B-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "B-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF", |
| "I-REF" |
| ], |
| "logits": [ |
| [ |
| -0.52291, |
| -4.8069, |
| -0.38009 |
| ], |
| [ |
| -0.31216, |
| -7.5731, |
| 1.51422 |
| ] |
| ] |
| } |
| } |
| ] |
| }, |
| "spec": { |
| "merge": "each token keeps the tag of the window where it is furthest from the window's edge; the first such window on a tie", |
| "reference_start": "first non-space character of every B-REF token, and of the text" |
| } |
| } |
|
|
