| { | |
| "model": "distilbert", | |
| "labels": [ | |
| "O", | |
| "B-PER", | |
| "I-PER", | |
| "B-ORG", | |
| "I-ORG", | |
| "B-LOC", | |
| "I-LOC", | |
| "B-MISC", | |
| "I-MISC" | |
| ], | |
| "length_policy": { | |
| "max_subwords": 256, | |
| "word_policy": "No recurrent word truncation; packed dynamic batches", | |
| "char_max": 61, | |
| "char_policy": "Train max; held-out/app longer chars clipped, count reported", | |
| "transformer_policy": "Greedy whole-word chunks; no overlap; label-independent; reconstruct before scoring", | |
| "app_recurrent_chunk_words": 256, | |
| "app_max_characters": 20000, | |
| "app_sentence_policy": "raw-v2: acronyms/titles and hyphenated words retained; format/control characters omitted from inference with original offsets preserved; standalone .!? and newlines split sentences" | |
| }, | |
| "selection": { | |
| "winner": "distilbert", | |
| "best_validation_f1": 0.9391641388608022, | |
| "eligible_within_0.001": [ | |
| "distilbert" | |
| ], | |
| "rule": "Strict validation micro F1; within absolute 0.001 prefer fewer trainable parameters, then measured validation seconds, then stable model name", | |
| "experiment_id": "8a1b26004cfaa53318dee7fbb6289ba971e3152c3aaaf7171b08086838f4aac2", | |
| "mode": "full", | |
| "checkpoint_manifests": { | |
| "distilbert": "b09581ff4d98bc34ec88817dc6357df0e38f57e33262f84ed7fca251888eca30", | |
| "lstm": "55fd4158f08af157b3466c47fc5c2ed50f295a29f7e6e390107d699146733e5a", | |
| "bilstm": "ed628fb397d73a3d130e5c91460d0b340402dcbea621cc3a7898e9d62606d04a", | |
| "bilstm_crf": "81979fc4eff8545034b555f1c4906f0e785bcdc70a257df0668a3a2cb85ebb5a" | |
| } | |
| }, | |
| "experiment_id": "8a1b26004cfaa53318dee7fbb6289ba971e3152c3aaaf7171b08086838f4aac2", | |
| "preprocessing": "Preserve casing/punctuation/word order; train-only vocab; lowercase only external GloVe lookup. No POS/chunk features.", | |
| "span_policy": "Strict IOB2; invalid I predictions ignored rather than repaired. Half-open original character offsets.", | |
| "app_tokenizer": "raw-v2: acronyms/titles and hyphenated words retained; format/control characters omitted from inference with original offsets preserved; standalone .!? and newlines split sentences", | |
| "benchmark_tokenization_note": "Raw-text regex is deterministic but differs from human CoNLL tokenization; exported token API preserves benchmark tokens exactly.", | |
| "embedding_provenance": { | |
| "num_records": 400000, | |
| "file_size": 69182535, | |
| "base_dataset": "Wikipedia 2014 + Gigaword 5 (6B tokens, uncased)", | |
| "reader_code": "https://github.com/RaRe-Technologies/gensim-data/releases/download/glove-wiki-gigaword-50/__init__.py", | |
| "license": "http://opendatacommons.org/licenses/pddl/", | |
| "parameters": { | |
| "dimension": 50 | |
| }, | |
| "description": "Pre-trained vectors based on Wikipedia 2014 + Gigaword, 5.6B tokens, 400K vocab, uncased (https://nlp.stanford.edu/projects/glove/).", | |
| "preprocessing": "Converted to w2v format with `python -m gensim.scripts.glove2word2vec -i <fname> -o glove-wiki-gigaword-50.txt`.", | |
| "read_more": [ | |
| "https://nlp.stanford.edu/projects/glove/", | |
| "https://nlp.stanford.edu/pubs/glove.pdf" | |
| ], | |
| "checksum": "c289bc5d7f2f02c6dc9f2f9b67641813", | |
| "file_name": "glove-wiki-gigaword-50.gz", | |
| "parts": 1 | |
| }, | |
| "training_mode": "full", | |
| "full_training_results": true, | |
| "inference_update": { | |
| "revision": "raw-v2", | |
| "date": "2026-11-10", | |
| "training_experiment_unchanged": true, | |
| "previous_app_archive": { | |
| "bytes": 242761064, | |
| "sha256": "079a09e2e8239c28f8818b414ba7810e8e51433016f7098d00f0dcae12c986b1" | |
| }, | |
| "source_sha256": { | |
| "common.py": "33b5d2a9dbb97e86ad46cf79c41433f8525ac76494ff9966f908f456d408e3a3", | |
| "inference.py": "0286bfa88af230330333402052bc6ddc58581f212fc039e78cc32889ba1e418f", | |
| "app.py": "09e7845cdfd24e85eca8cb34a5f043e8fee7a047866fe892352944fabe4acc30", | |
| "models.py": "05519b51d548b7cdca9df3248d51d44e0916937f0711fb4ce3913bc0a6a25cb0" | |
| }, | |
| "verification": "All 6 raw-text and offline CPU checkpoint regression tests passed, including former crashes, full acronym/hyphen boundaries, offsets and long input. Updated Kaggle training has not been rerun." | |
| } | |
| } |