cstr/punctuate-all-GGUF

Punctuate-All (GGUF)

GGUF conversion of kredor/punctuate-all for use with CrispASR.

Adds punctuation to unpunctuated ASR output. 12 languages with ASCII punctuation output. Smaller and faster alternative to fullstop-punc-multilang (base vs large).

Model Details

• Architecture: XLM-RoBERTa-base — 12L, d=768, 12 heads, d_ffn=3072, GELU
• Parameters: ~278M
• Classifier: Linear(768, 6) — 6 punctuation classes
• Labels: none, . (period), , (comma), ? (question), - (dash), : (colon)
• Vocabulary: SentencePiece (250,002 tokens)
• Max sequence: 512 tokens (auto-chunked)
• Languages: en, de, fr, es, bg, it, pl, nl, cs, pt, sk, sl
• License: MIT

Usage with CrispASR

crispasr --backend wav2vec2 -m wav2vec2.gguf --punc-model punctuate-all -f audio.wav

Available Files

File	Quant	Size	Description
punctuate-all-q4_k.gguf	Q4_K	162 MB	Recommended. Rebuilt 2026-08-25
punctuate-all-f16.gguf	F16	946 MB	Half precision. Rebuilt 2026-08-25
punctuate-all-orig-q4_k.gguf	Q4_K	161 MB	The original conversion, kept verbatim
punctuate-all-orig-f16.gguf	F16	945 MB	The original conversion, kept verbatim

What changed on 2026-08-25

The canonical files were rebuilt. The originals are preserved unchanged as punctuate-all-orig-* — nothing was deleted.

One real fix. The original conversion shipped tokenizer.ggml.tokens but no tokenizer.ggml.scores, so the runtime fell back to greedy longest-match. XLM-R's SentencePiece model is Unigram, where greedy is not an approximation but the wrong algorithm: fox has no ▁fox piece, so Viterbi produces ▁ + fox (ids 6, 147797) while greedy takes the longest prefix ▁fo and is left with x (5775, 425). Different ids, different embeddings, silently. On public-domain prose the original matched HuggingFace's tokenizer on 0 of 7 segments; the rebuild matches on 6 of 7.

One thing deliberately kept. kredor/punctuate-all zeroes 9531 token embedding rows — four contiguous ranges (4086–5449, 6816–9545, 10912–12276, 51895–55966), which look like pruned token ranges for languages it does not serve. The original conversion happened to retain xlm-roberta-base's vectors there. That turns out to help, so the rebuild keeps it deliberately rather than by accident, via --restore-zeroed-embeddings-from xlm-roberta-base.

Measured

120 sentences of public-domain prose (English, German, French — 40 each, 2349 words, 350 marks), scoring restored punctuation against the original editors'. markF1 is exact-mark agreement; bndF1 asks only whether a sentence ended there, which is far less sensitive to house style.

artifact	markF1	bndF1	per-word exact
punctuate-all-q4_k (rebuild)	0.771	0.922	0.944
punctuate-all-orig-q4_k	0.762	0.922	0.941
faithful rebuild with kredor's zeroed rows	0.713	0.884	0.931

Paired bootstrap, 2000 resamples, 95% interval on the markF1 difference:

rebuild vs orig                      +0.0092  [-0.0174, +0.0344]  not distinguishable
rebuild vs kredor's zeroed rows      +0.0542  [+0.0245, +0.0838]  significant

So the rebuild is at least as good as the original on punctuation quality and additionally tokenizes correctly. Restoring the zeroed embeddings is a real, significant gain — those zero rows cost quality.

⚠ Note for anyone running a parity check

No file in this repo will match a reference dumped from kredor/punctuate-all via transformers on text that uses the 9531 zeroed rows — not the rebuild and not the originals, for the same reason in both cases. transformers loads the zeros; every file here carries base vectors there.

That is by design, not a defect, and it is why the rebuild scores where it does. If you need blueprint-exact behaviour, convert kredor/punctuate-all yourself without --restore-zeroed-embeddings-from and expect the third row of the table above (markF1 0.713).

Original Model

• Source: kredor/punctuate-all
• Training: WMT/Europarl dataset
• Conversion: models/convert-fullstop-punc-to-gguf.py in CrispASR