Hugging Face's logo

juanquivilla
/
sotto-cleanup-lfm25-350m-mlx-5bit

Text Generation
MLX
Safetensors
English
lfm2
speech-to-text
transcript-cleanup
text-correction
asr-post-processing
LFM
LiquidAI
mlx-5bit
conversational
5-bit
Model card Files
xet
 Community
juanquivilla commited on
Commit
1b04172
·
verified ·
1 Parent(s): 8152256

soup_30: composite=89.45 — see model card for benchmark deltas vs v45

Browse files
Files changed (2) hide show
  1. README.md +4 -4
  2. model.safetensors +1 -1
README.md CHANGED
@@ -15,7 +15,7 @@ tags:
15
  pipeline_tag: text-generation
16
  ---
17
 
18
- # SottoASR Transcript Cleanup — LFM2.5-350M MLX 5-bit (v55)
19
 
20
  [sottoasr.app](https://sottoasr.app) · [Full precision (bf16)](https://huggingface.co/juanquivilla/sotto-cleanup-lfm25-350m) · [MLX 4-bit (smaller)](https://huggingface.co/juanquivilla/sotto-cleanup-lfm25-350m-mlx-4bit)
21
 
@@ -23,13 +23,13 @@ pipeline_tag: text-generation
23
 
24
  MLX 5-bit affine quantization of [juanquivilla/sotto-cleanup-lfm25-350m](https://huggingface.co/juanquivilla/sotto-cleanup-lfm25-350m). Recommended for Apple Silicon — best size/quality trade-off.
25
 
26
- ## What's new in v55
27
 
28
- v55 extends v45 with targeted training data for five failure modes (multi-number sentences,
29
  year-context drift, disconnected number lists, within-input duplicates, long-form preservation),
30
  each generated programmatically and audited with a Qwen3.6-27B judge.
31
 
32
- | Metric | v45 | **v55** |
33
  |---|---:|---:|
34
  | Number accuracy | 95.9% | **96.5%** |
35
  | Adversarial benchmark (greedy) | 76% | **86%** |
 
15
  pipeline_tag: text-generation
16
  ---
17
 
18
+ # SottoASR Transcript Cleanup — LFM2.5-350M MLX 5-bit (soup_30)
19
 
20
  [sottoasr.app](https://sottoasr.app) · [Full precision (bf16)](https://huggingface.co/juanquivilla/sotto-cleanup-lfm25-350m) · [MLX 4-bit (smaller)](https://huggingface.co/juanquivilla/sotto-cleanup-lfm25-350m-mlx-4bit)
21
 
 
23
 
24
  MLX 5-bit affine quantization of [juanquivilla/sotto-cleanup-lfm25-350m](https://huggingface.co/juanquivilla/sotto-cleanup-lfm25-350m). Recommended for Apple Silicon — best size/quality trade-off.
25
 
26
+ ## What's new in soup_30
27
 
28
+ soup_30 extends v45 with targeted training data for five failure modes (multi-number sentences,
29
  year-context drift, disconnected number lists, within-input duplicates, long-form preservation),
30
  each generated programmatically and audited with a Qwen3.6-27B judge.
31
 
32
+ | Metric | v45 | **soup_30** |
33
  |---|---:|---:|
34
  | Number accuracy | 95.9% | **96.5%** |
35
  | Adversarial benchmark (greedy) | 76% | **86%** |
model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f98aef88fca96c05b54693f7bc9fde66e00691068d35b8b6f202cb501b7d29b9
3
  size 243830312
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e9f5057c16184b904ce4d7e06eba5b54cd0cb46117141ebeb14db3c5ad48b61
3
  size 243830312