jiangzhexin/bgm-tagger-bilstm

BGM Tagger BiLSTM

A small character-level bidirectional LSTM sequence tagger for parsing anime release filenames into structured entities. Given a raw filename, it tags every character with a BIO label and reconstructs an XML fragment such as:

[LoliHouse] 婚戒物语 / Kekkon Yubiwa Monogatari - 01 [WebRip 1080p HEVC-10bit AAC][简繁内封字幕]

[<ReleaseGroup>LoliHouse</ReleaseGroup>] <TitleEntity>婚戒物语</TitleEntity> / <TitleEntity>Kekkon Yubiwa Monogatari</TitleEntity> - <Episode>01</Episode> [<SourceEntity>WebRip</SourceEntity> <Resolution>1080p</Resolution> <VideoCodec>HEVC</VideoCodec>-<VideoDepth>10bit</VideoDepth> <AudioCodec>AAC</AudioCodec>][<SubtitleLanguage>简繁</SubtitleLanguage><SubtitleFormat>内封</SubtitleFormat>字幕]

Quick start

uv add bgm-tagger
# or
pip install bgm-tagger

from bgm_tagger import BiLSTM

model = BiLSTM.from_pretrained("jiangzhexin/bgm-tagger-bilstm")
model.eval()

# Move to GPU if available
model.to("cuda")

text = "【喵萌奶茶屋】★01月新番★[迷宫饭 / Dungeon Meshi / Delicious in Dungeon][12][1080p][简日双语][招募翻译时轴]"

# One XML fragment per input filename
print(*model.predict(text), sep="\n")

# Or raw BIO tag sequences
print(*model.predict(text, return_type="bio"), sep="\n")

The output for the example above is:

【<ReleaseGroup>喵萌奶茶屋</ReleaseGroup>】★01月新番★[<TitleEntity>迷宫饭</TitleEntity> / <TitleEntity>Dungeon Meshi</TitleEntity> / <TitleEntity>Delicious in Dungeon</TitleEntity>][<Episode>12</Episode>][<Resolution>1080p</Resolution>][<SubtitleLanguage>简日</SubtitleLanguage>双语][招募翻译时轴]

Model architecture

A character-level sequence tagger with no subword tokenization or pretrained embeddings. Each filename is encoded as a sequence of character ids, processed by a two-layer bidirectional LSTM, and decoded with BIO transition constraints.

Component	Configuration
Tokenizer	Character vocabulary (1,750 chars + <PAD> + <UNK>)
Embedding	nn.Embedding(1752, 256), padding_idx=0
Dropout	0.25
Encoder	2-layer bidirectional LSTM, hidden size 256, dropout 0.25
Classifier	Linear(512, 25) over the concatenated forward/backward states
Decoding	Constrained Viterbi-style decoding with hard BIO transition rules
Parameters	~3.09M (all trainable)

Labels

The model predicts 25 labels: B-/I- prefixes for 12 entity types plus O.

Entity	Example
TitleEntity	迷宫饭, Dungeon Meshi
Episode	01, 01-13
Season	Season 2, S1-S4
ReleaseGroup	LoliHouse, VCB-Studio, 喵萌奶茶屋
SourceEntity	WebRip, BDRip, WEB-DL, Baha
Resolution	1080p, 720p
VideoCodec	HEVC, AVC
AudioCodec	AAC, FLAC
VideoDepth	10bit, 8bit
SubtitleLanguage	简日, 简繁, CHT
SubtitleFormat	内封, ass
Version	v2

Training procedure

The model was trained on jiangzhexin/bgm-bio-xml, a dataset of 10,000 anime release filenames annotated with BIO/XML entities.

Hyperparameter	Value
Train / validation / test split	7,225 / 1,275 / 1,500 (last 15% held out as test; the rest shuffled 85/15)
Optimizer	AdamW (lr=0.001, weight_decay=1e-4)
Learning rate scheduler	ReduceLROnPlateau (factor=0.5, patience=2, min_lr=5e-5)
Batch size	128
Max epochs	40 (early stopping with patience=8, min_delta=1e-4)
Seed	42

Training stopped early after 18 epochs. The checkpoint restored for evaluation is the one with the best validation entity micro-F1.

Evaluation

Metrics are computed on the held-out test split (1,500 examples).

Metric	Value
Test loss	0.0054
Entity micro-F1	0.9987
Exact-match accuracy	0.984

• Entity micro-F1 counts exact entity spans (type + boundaries) across all sequences.
• Exact-match accuracy is the fraction of filenames whose predicted entity set matches the gold entity set exactly.

Best validation entity micro-F1 during training: 0.9988 (epoch 17).

Limitations

• The model is trained on nyaa.si anime release filenames and may not generalize to other naming conventions.
• It tags characters, so unusual symbols outside the training vocabulary are mapped to <UNK>.
• Decoding uses hard BIO constraints rather than a learned CRF, so label transition probabilities are not modeled.