Vektor MLX Models
Collection
6 items • Updated
How to use Nocti-Dev/nb-whisper-medium-mlx with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download Nocti-Dev/nb-whisper-medium-mlx --local-dir nb-whisper-medium-mlx
MLX conversion of NbAiLab/nb-whisper-medium at commit 0ed074d5985bd56ca4140159a9dbffbc3fb5117e, made with Prism. Self-contained: weights, config and tokenizer are all in this repo, so it loads offline.
| Source | NbAiLab/nb-whisper-medium @ 0ed074d5985b |
| Recipe | whisper-hf v1 |
| dtype | float16 |
| Quantization | none |
| Licence | apache-2.0 (from NbAiLab/nb-whisper-medium) |
| Worker | prism-worker 0.1.0, Python 3.14.10, mlx 0.32.4, numpy 2.5.4, huggingface_hub 2.3.0, hf-xet 1.7.0, torch 2.14.1, transformers 5.21.0 |
| Worker lockfile | 6f31327071399b5c (sha256) |
The unconverted model (NbAiLab/nb-whisper-medium@0ed074d, transformers 5.21.0 / torch 2.14.1) and this conversion (mlx-audio-swift@dbe5eaa) transcribed the same short Norwegian test clip: passed, the transcripts differ by a word error rate of 0.00 (max 0.10).
| Word error rate | |
|---|---|
| Converted vs unconverted (verdict) | 0.00 |
| Converted vs script | 0.49 |
| Unconverted vs script | 0.49 |
| File | Size | sha256 |
|---|---|---|
added_tokens.json |
34,604 | 9715fd2243b6f06a5858b5e32950d2853f73dd5bc201aafcf76f5082a2d8acd1 |
config.json |
4,504 | 7a6279ebab765c74d2d81454cdaee3de3c5b660cf8efd4eb2d4c252004a771b7 |
generation_config.json |
3,802 | 45895ed5d67c8453df306a22c8b2d357a63c55965cacc3bbbbc6928d38f9f1d2 |
merges.txt |
493,864 | 78f27f801feb4283ae969d4cadacb28fc051b655647402b1a252c7079ea11c46 |
model.safetensors |
1,527,827,567 | 8da0e24e70d388b7dbcfb5f6b3538254eb7fe7166fbc2563372b9bf2bf7cfd74 |
normalizer.json |
52,666 | bf1c507dc8724ca9cf9903640dacfb69dae2f00edee4f21ceba106a7392f26dd |
preprocessor_config.json |
339 | a6a76d28c93edb273669eb9e0b0636a2bddbb1272c3261e47b7ca6dfdbac1b8d |
special_tokens_map.json |
2,190 | 4679897a50cc1aea98b7d63b8c928b407f252683d3248add11d091e66cf0d5f1 |
tokenizer.json |
2,481,189 | 87b38f7a654dbbda293bcdda048d2fb18af241fdd41c62496f8e732209610ec1 |
tokenizer_config.json |
282,699 | 348e80c42b59d4e005fdc7a68ad0244ed339e30eaaabcf805e8e28ef73e7d1e8 |
vocab.json |
835,550 | 8f680bba319e01a653d2e8a5dbc17a9157179e0576e6ce74ce0c06356c6e24f9 |
Full provenance is in prism.json.
Quantized