README.md

metadata

pretty_name: AnimeName
license: other
language:
  - ja
  - zh
  - en
task_categories:
  - token-classification

AnimeName

Weak-labeled anime filename datasets for training filename parsers.

Contents

• dmhy_weak.jsonl: weak labels exported from the local DMHY SQLite crawl snapshot
• dmhy_weak_char.jsonl: the same DMHY weak labels projected to one Unicode code point per token for character-level training
• mixed_train.jsonl: 100K synthetic templates mixed with DMHY weak labels
• ab_mix_100k.jsonl: balanced 50K synthetic + 50K DMHY A/B comparison subset
• *.manifest.json: export waterlines, sample counts, and label stats
• vocab.json: regex tokenizer vocab for the main parser snapshot
• vocab.char.json: full DMHY character tokenizer vocab
• vocab.char.ab100k.json: character tokenizer vocab for the A/B experiment

Snapshot

• Export source: D:\WorkSpace\Python\dmhy-parser\dmhy_anime.db
• DB max files.id: 1675184
• Last exported files.id: 1675184
• Next incremental export: --min-id 1675185
• Weak-labeled samples: 632002
• Character weak-labeled samples: 632002
• Full mixed training set: 732002
• A/B balanced subset: 100000

Character-token DMHY Export

dmhy_weak_char.jsonl was generated directly from dmhy_weak.jsonl by splitting every labeled token into characters and projecting BIO labels:

• B-X: first character keeps B-X, remaining characters become I-X
• I-X: every character stays I-X
• O: every character stays O

The full DMHY weak dataset contains 6195 unique characters. With the four special tokens ([PAD], [UNK], [CLS], [SEP]), vocab.char.json has 6199 entries and covers 100% of character-token occurrences. The sequence length distribution is p50=60, p90=83, p95=91, p99=107, max=170 characters.

v3 Export Notes (current)

• Fixed ordinal season patterns: 3rd Season, 4th Season etc. now correctly captured as single tokens and labeled B-SEASON (was broken: tokenizer regex \d+[sn][dt] only matched st/nd, missed rd/th).
• Fixed 第N部 (Chinese "Part N") season detection.
• Bracket expansion splits mixed-content brackets (e.g. [剑来 第2季]) so that 第2季 is correctly labeled as B-SEASON.
• Season labels total: 66,497 (+47.9% from original 44,964).
• Wikipedia cross-reference verified — 3 multi-season anime checked: Wikipedia season data was retrieved and compared against the weak labels. All 3 shows' season structures (S1 episode ranges, S2+, split-cours) were verified to match the labels produced by the heuristic pipeline.
  • Jujutsu Kaisen: S1(24eps 2020), S2(23eps 2023), S3(12eps 2026)
  • Mushoku Tensei: S1(23eps 2021), S2(25eps 2023-24)
  • Re:Zero: S1(25eps 2016), S2(25eps 2020-21), S3(16eps 2024-25) All explicit season markers (S1, S2, Season N, 第N季) in filenames for these anime match their Wikipedia season numbers. First seasons often omit markers (fansub convention), which the inference logic handles via proximity detection.
• A small number of token sequences now show I-GROUP following an O separator (by design: assign_bio allows entity spans across separators).

Known Limitations

• Zombie I- labels: config.json defines I-EPISODE, I-SEASON, I-SPECIAL, I-RESOLUTION, and I-SOURCE labels, but they never appear in training data because the regex tokenizer keeps these entity types as single tokens. These 5 unused classifier heads out of 15 total are harmless (~0.06% of parameters).

Notes

The JSONL files are raw dataset snapshots. Model checkpoints are intentionally not included here.