OpenASR commited on
Commit
a96d9ef
·
verified ·
1 Parent(s): 1cbb691

docs: refresh card for fp16-only ship

Browse files
Files changed (1) hide show
  1. README.md +4 -7
README.md CHANGED
@@ -51,12 +51,9 @@ openasr transcribe meeting.wav --model xasr-zh-en --diarize --format srt
51
  | Quant | File (`.oasr`) | Size |
52
  |:------|:---------------|-----:|
53
  | fp16 | `redimnet2-b6-cn-fp16.oasr` | 28 MB |
54
- | q8_0 | `redimnet2-b6-cn-q8_0.oasr` | 27 MB |
55
- | f32 | `redimnet2-b6-cn-f32.oasr` | 50 MB |
56
 
57
- <sub>Single raw-**f32** build: the pure-Rust forward pass consumes f32 directly and the
58
- parity gates assert bit-exact outputs vs the upstream weights, so no integer
59
- quantization is produced.</sub>
60
 
61
  ## 🧠 About ReDimNet2-B6 Speaker Embedder (CN-enhanced)
62
 
@@ -79,8 +76,8 @@ openasr model-pack import redimnet2 <src>.safetensors <out>.oasr \
79
  --package-id redimnet2-b6-cn
80
  ```
81
 
82
- The `.oasr` container is GGUF-backed; every tensor is stored as raw f32 so the
83
- pack round-trips bit-identically against the source weights.
84
 
85
  ## ⚖️ License
86
 
 
51
  | Quant | File (`.oasr`) | Size |
52
  |:------|:---------------|-----:|
53
  | fp16 | `redimnet2-b6-cn-fp16.oasr` | 28 MB |
 
 
54
 
55
+ <sub>Single **fp16** build: projection weights ship as fp16; norms/biases and other
56
+ parity-sensitive tensors stay f32 inside the pack. No extra public quant tiers.</sub>
 
57
 
58
  ## 🧠 About ReDimNet2-B6 Speaker Embedder (CN-enhanced)
59
 
 
76
  --package-id redimnet2-b6-cn
77
  ```
78
 
79
+ The `.oasr` container is GGUF-backed; projection weights are stored as fp16 while
80
+ norms/biases and other parity-sensitive tensors remain f32.
81
 
82
  ## ⚖️ License
83