cstr/fireredpunc-GGUF

FireRedPunc (GGUF)

GGUF conversion of FireRedTeam/FireRedPunc for use with CrispASR.

Adds punctuation to unpunctuated ASR output using a BERT-based token classifier.

Model Details

• Architecture: BERT-base (LERT) — 12L, d=768, 12 heads, d_ffn=3072, GELU
• Parameters: ~102M
• Classifier: Linear(768, 5) — 5 punctuation classes
• Labels: space, ，(comma), 。(period), ？(question), ！(exclamation)
• Vocabulary: chinese-bert-wwm-ext (21,128 tokens, WordPiece)
• Max sequence: 512 tokens (longer texts are automatically chunked)
• Languages: Chinese + English
• License: Apache 2.0

Usage with CrispASR

# Add punctuation to any ASR backend's output
crispasr --backend wav2vec2 -m wav2vec2.gguf --punc-model fireredpunc-q8_0.gguf -f audio.wav

# Works with all backends
crispasr --backend omniasr -m omniasr-ctc-1b-q4_k.gguf --punc-model fireredpunc-q8_0.gguf -f audio.wav

Available Files

File	Quant	Size	Description
fireredpunc.gguf	F16	195 MB	Full precision
fireredpunc-q8_0.gguf	Q8_0	104 MB	Recommended — lossless quality
fireredpunc-q4_k.gguf	Q4_K	56 MB	Smaller but may miss some punctuation

Recommendation: Use Q8_0. Token classification is more sensitive to quantization than language modeling — Q4_K occasionally drops commas that Q8_0 and F16 correctly predict.

Example

Input (from CTC ASR):

and so my fellow americans ask not what your country can do for you ask what you can do for your country

Output (with FireRedPunc):

And so my fellow americans， ask not what your country can do for you， ask what you can do for your country？

Conversion

python models/convert-fireredpunc-to-gguf.py \
  --input /path/to/FireRedPunc \
  --output fireredpunc.gguf

Original Model

• Source: FireRedTeam/FireRedPunc
• Paper: FireRedASR2S
• Base: LERT (Linguistically-motivated BERT)
• Training: 18.57B Chinese characters + 2.20B English words

Provenance and EU AI Act Art. 53 note

• Upstream model: FireRedTeam/FireRedPunc — published by FireRedTeam.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.