| { |
| "id": "bashkir_diacritics_restorer", |
| "version": "v1.2", |
| "type": "sequence_tagging_restorer", |
| "language": "ba", |
| "status": "release_candidate", |
| "license": "cc-by-4.0", |
| "files": ["model.onnx", "vocab.json", "substitution_map.json", "config.json", "runtime.py"], |
| "architecture": "ConvBiGRU (Conv1D 3,5 + 2-layer BiGRU 192)", |
| "parameters": 1184032, |
| "metrics": { |
| "original_heldout": {"sentence_exact_match": 0.9342, "spec_letter_f1": 0.9967}, |
| "historical_safety_stress_n_1000": { |
| "note": "Earlier experimental-runtime measurements; the exact evaluation protocol was not recorded and the public-only 87.8% figure is not reproducible on the current test file.", |
| "with_optional_kenlm": {"restoration_exact_match": 0.923, "clean_input_untouched_pct": 98.5}, |
| "default_no_kenlm": {"restoration_exact_match": 0.891, "clean_input_untouched_pct": 97.8}, |
| "neural_only_no_lexicon_no_kenlm": {"restoration_exact_match": 0.878, "clean_input_untouched_pct": 97.6} |
| }, |
| "recheck_2026_09_30": { |
| "test_file": "experiments/diacritics_restorer/data/test.parquet", |
| "test_file_sha256": "72230475aa015c04a6ae0cf8bd17aa9420ededa75ba9d6447b7d94a4ffb54088", |
| "selection": "first 1000 rows in file order", |
| "threshold": 0.40, |
| "public_runtime_without_lexicon": {"restoration_exact_match": 0.836, "clean_input_untouched_pct": 97.6}, |
| "public_runtime_with_local_optional_lexicon": {"restoration_exact_match": 0.888, "clean_input_untouched_pct": 97.6}, |
| "experimental_runtime_without_lexicon": {"restoration_exact_match": 0.838, "clean_input_untouched_pct": 97.6}, |
| "experimental_runtime_with_local_lexicon": {"restoration_exact_match": 0.891, "clean_input_untouched_pct": 97.8} |
| } |
| }, |
| "runtime": {"safe_mode": true, "kenlm": "optional; excluded from release"}, |
| "source_policy": "Model artifacts and derived substitution rules only; training text and derived lexical data are not redistributed." |
| } |
|
|
