Fleet-Clean-0.36B-GGUF
GGUF files for Fleet-Clean-0.36B: speech in, clean text out, with no Cactus or PyTorch needed.
| File | What | Runs in |
|---|---|---|
whistle-fleet2-F16.gguf (recommended), -F32.gguf |
Fleet-2 speech-to-text (Cactus Whistle architecture) | whistle/whistle-cli (C++, in this repo) |
fleet-clean-tagger-Q8_0.gguf / -F16.gguf |
the cleanup tagger's encoder (modern-bert) |
llama.cpp llama-server --embeddings --pooling none |
fleet-clean-tagger-head.gguf |
the 768→10 head and its labels, read by fleet_clean_gguf.py |
numpy |
Usage
hf download Lucasgitguy-co/Fleet-Clean-0.36B-GGUF --local-dir fleet-clean-gguf && cd fleet-clean-gguf
sh whistle/build.sh # whistle-cli: Accelerate on macOS, OpenBLAS or plain C++ elsewhere
pip install numpy gguf # plus llama.cpp's llama-server on PATH (a build with modern-bert support)
python fleet_clean_gguf.py clip.wav # 16 kHz WAV -> clean text
python fleet_clean_gguf.py --text "so, um, I I think we should, you know, go"
whistle/whistle-cli -m whistle-fleet2-F16.gguf -f clip.wav --json # speech only
Important: GGUF speech is less accurate than whistle.cact
No existing GGUF runtime supports Whistle, so whistle-cli is a new C++ port of Cactus's open JAX model. Cactus's own .cact files run in a closed native engine, and on the same weights that engine scores better:
| Language | Clips | whistle.cact (native engine) |
whistle-fleet2-F32/F16.gguf (open model) |
|---|---|---|---|
| English | 100 | 14.5% | 17.6% |
| German | 100 | 22.0% | 25.1% |
| French | 30 | 27.4% | 30.7% |
| Spanish | 30 | 11.3% | 12.5% |
| Italian | 30 | 18.6% | 19.3% |
| Dutch | 30 | 32.5% | 36.6% |
| Polish | 30 | 43.2% | 48.6% |
| Korean | 30 | 34.4% CER | 45.3% CER |
The weights in these GGUF files are exactly Fleet-2's: they were dequantised from whistle.cact, and a re-export check confirmed the tensor mapping. The gap comes from the engines. Beam search, padding, silence trimming and 8-bit activations were all tested and don't explain it. For the best accuracy, use whistle.cact with Fleet-Clean-0.36B.
Parity checks
parity/ has every number.
| File | Transcripts identical to the open JAX model | Shipped |
|---|---|---|
whistle-fleet2-F32.gguf |
380/380 | yes |
whistle-fleet2-F16.gguf |
378/380 | yes |
whistle-fleet2-Q8_0.gguf |
344/380 | no (fails the gate) |
| Tagger file | Word tags identical to PyTorch | Token-id mismatches | Shipped |
|---|---|---|---|
fleet-clean-tagger-F16.gguf |
99.99% | 0 | yes |
fleet-clean-tagger-Q8_0.gguf |
99.82% | 0 | yes |
- The Whistle Q8_0 file failed its gate (at least 95% of transcripts identical to the open model, and WER no more than 0.5 points worse), so it isn't included.
- The tagger GGUF uses the original jhu-clsp/mmBERT-base tokenizer files. They have the same vocabulary and merges as Laya's, which transformers 5 re-saved in a form llama.cpp rejects. Token ids were checked against the HF tokenizer on every evaluation text.
- In F16, a few tokens with outlier activations ("I", ".") reach about 0.98 cosine against PyTorch on every backend, while the mean is ≥ 0.9999. The word tags are what matter, and they are compared above.
whistle-clireads GGUF itself (F32 / F16 / Q8_0), uses float32 throughout and decodes greedily, the same way as Cactus's reference decoder. It handles 16-bit, 24-bit and 32-bit PCM or float WAV, up to 30 s.
License
Apache-2.0. whistle/whistle.cpp is a port of Apache-2.0 Cactus code. See NOTICE.