notforu777 commited on
Commit
feed993
·
verified ·
1 Parent(s): 82d6923

Card: mmproj F16 is 0.62 GB

Browse files
Files changed (1) hide show
  1. README.md +3 -3
README.md CHANGED
@@ -59,14 +59,14 @@ You need **one backbone file + one mmproj**. The mmproj holds the codebook heads
59
  | mmproj | GB on disk | CER EN / DE / ZH (BF16 backbone) | voice similarity vs F32 (EN / DE / ZH) |
60
  |---|---|---|---|
61
  | `mmproj-ZONOS3-Mini-F32.gguf` | 1.07 | 0.06% / 0.00% / 0.91% | reference |
62
- | **`mmproj-ZONOS3-Mini-F16.gguf`** | **0.54** | 0.03% / 0.00% / 1.08% | +0.001 / −0.001 / −0.003 |
63
 
64
  **Bold numbers** are differences whose 95% interval excludes zero. All others are within the sampling noise of BF16.
65
 
66
- - **Recommended: AD-Q4_K_M + mmproj F16**, about 1.05 GB in total. In this test it is within noise of BF16 everywhere except a small drop in German voice similarity.
67
  - **Smallest: AD-Q3_K_M** (0.41 GB). Its losses are small but measurable: Chinese CER +0.7 pt, voice similarity −0.004 to −0.007.
68
  - The **stock Q4_K_M** preset is the same size, and it loses more on Chinese (+1.8 pt CER, significant). The AD file keeps everything except the experts at Q8_0 and moves the experts' `ffn_down` one step up.
69
- - The mmproj **F16 is within noise of F32** on these metrics. The PR suggests F32, so it is kept here too.
70
 
71
  Samples (AD-Q4_K_M + mmproj F32, seed 1):
72
 
 
59
  | mmproj | GB on disk | CER EN / DE / ZH (BF16 backbone) | voice similarity vs F32 (EN / DE / ZH) |
60
  |---|---|---|---|
61
  | `mmproj-ZONOS3-Mini-F32.gguf` | 1.07 | 0.06% / 0.00% / 0.91% | reference |
62
+ | **`mmproj-ZONOS3-Mini-F16.gguf`** | **0.62** | 0.03% / 0.00% / 1.08% | +0.001 / −0.001 / −0.003 |
63
 
64
  **Bold numbers** are differences whose 95% interval excludes zero. All others are within the sampling noise of BF16.
65
 
66
+ - **Recommended: AD-Q4_K_M + mmproj F16**, about 1.13 GB in total. In this test it is within noise of BF16 everywhere except a small drop in German voice similarity.
67
  - **Smallest: AD-Q3_K_M** (0.41 GB). Its losses are small but measurable: Chinese CER +0.7 pt, voice similarity −0.004 to −0.007.
68
  - The **stock Q4_K_M** preset is the same size, and it loses more on Chinese (+1.8 pt CER, significant). The AD file keeps everything except the experts at Q8_0 and moves the experts' `ffn_down` one step up.
69
+ - The mmproj **F16 is within noise of F32** on these metrics and is 0.45 GB smaller (some of its tensors stay F32). The PR suggests F32, so it is kept here too.
70
 
71
  Samples (AD-Q4_K_M + mmproj F32, seed 1):
72