hletrd commited on
Commit
f181f44
·
verified ·
1 Parent(s): 41619c5

Card: publish corrected accuracy and metric charts

Browse files
.gitattributes CHANGED
@@ -35,3 +35,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  accuracy-by-size.png filter=lfs diff=lfs merge=lfs -text
37
  accuracy-by-size-dark.png filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  accuracy-by-size.png filter=lfs diff=lfs merge=lfs -text
37
  accuracy-by-size-dark.png filter=lfs diff=lfs merge=lfs -text
38
+ accuracy-kl-dark.png filter=lfs diff=lfs merge=lfs -text
39
+ accuracy-kl.png filter=lfs diff=lfs merge=lfs -text
40
+ accuracy-top1-dark.png filter=lfs diff=lfs merge=lfs -text
41
+ accuracy-top1.png filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -66,12 +66,17 @@ Mean KL divergence from BF16 (lower is better) and top-1 agreement with BF16.
66
 
67
  | Quant | Size | WikiText-2 KL | top-1 | Chat KL | top-1 | Held-out KL | top-1 |
68
  |---|---:|---:|---:|---:|---:|---:|---:|
69
- | **MLX 4-bit** | 447.58 GB | **0.128577** | **90.588%** | | | | |
70
- | **MLX 8-bit** | 819.63 GB | | | | | | |
71
 
72
  <picture>
73
- <source media="(prefers-color-scheme: dark)" srcset="accuracy-by-size-dark.png">
74
- <img src="accuracy-by-size.png" alt="Mean KL and top-1 by file size">
 
 
 
 
 
75
  </picture>
76
 
77
  ## License and notices
 
66
 
67
  | Quant | Size | WikiText-2 KL | top-1 | Chat KL | top-1 | Held-out KL | top-1 |
68
  |---|---:|---:|---:|---:|---:|---:|---:|
69
+ | **MLX 4-bit** | **447.58 GB** | **0.142528** | **88.549%** | **0.193011** | **90.126%** | **0.066185** | **92.363%** |
70
+ | **MLX 8-bit** | **819.63 GB** | | | | | | |
71
 
72
  <picture>
73
+ <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/Xylolabs/Hy4-MLX-4bit/resolve/main/accuracy-kl-dark.png">
74
+ <img src="https://huggingface.co/Xylolabs/Hy4-MLX-4bit/resolve/main/accuracy-kl.png" alt="Mean KL divergence by file size">
75
+ </picture>
76
+
77
+ <picture>
78
+ <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/Xylolabs/Hy4-MLX-4bit/resolve/main/accuracy-top1-dark.png">
79
+ <img src="https://huggingface.co/Xylolabs/Hy4-MLX-4bit/resolve/main/accuracy-top1.png" alt="Top-1 agreement by file size">
80
  </picture>
81
 
82
  ## License and notices
accuracy-by-size.csv CHANGED
@@ -1,3 +1,7 @@
1
- provider,quant,bytes,kl,top1,text,context,chunk_indices,chunk_count,scored_positions,scored_predictions,reference,reference_revision,software,batch,measurement_scope,source,size_source,measurement_note
2
- Xylolabs,MLX 4-bit,447576537740,0.128577,90.588,WikiText-2 raw test,512,0,1,256-510,255,tencent/Hy4 BF16 GGUF with native FP8 Engram,c10c0de5b6f850703fb0a90e65d1c32b558b9167,MLX 0.32.4; mlx-lm 0.32.2; bf16 activations; fp32 LM head,1,single valid chunk; full corrected aggregates pending,https://huggingface.co/Xylolabs/Hy4-MLX-4bit/blob/ad51910f4652b0511bd02b413e50523b20dc4314/README.md,pinned live Hub listing,existing published measurement
3
- Xylolabs,MLX 8-bit,819628080300,,,WikiText-2 raw test,512,0,1,256-510,255,tencent/Hy4 BF16 GGUF with native FP8 Engram,c10c0de5b6f850703fb0a90e65d1c32b558b9167,MLX 0.32.4; mlx-lm 0.32.2; bf16 activations; fp32 LM head,1,single valid chunk; full corrected aggregates pending,https://huggingface.co/Xylolabs/Hy4-MLX-4bit/blob/ad51910f4652b0511bd02b413e50523b20dc4314/README.md,pinned live Hub listing,pending
 
 
 
 
 
1
+ provider,quant,bytes,kl,top1,text,context,chunk_indices,chunk_count,scored_positions,scored_predictions,ppl,reference,reference_revision,reference_batch,reference_ubatch,reference_software,software,batch,input_sha256,result_log_sha256,weight_repo,weight_revision,source,size_source,measurement_note,chart_selected
2
+ Xylolabs,MLX 4-bit,447576537740,0.1425278112316929,88.54901960784314,WikiText-2 raw test,512,0-79,80,256-510,20400,3.3099513725070513,tencent/Hy4 BF16 GGUF with native FP8 Engram; checked redo bases,c10c0de5b6f850703fb0a90e65d1c32b558b9167,12288,12288,llama.cpp; checked BIG_BASES2 redo completed 2026-11-21,MLX 0.32.4; mlx-lm 0.32.2; CUDA 13; bf16 activations; fp32 LM head; zero cache; sequential layer loading,1,173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08,3766cffb20a064f19e033e41bd9617fe9157d12cebaff439ca5055f1b96731d1,Xylolabs/Hy4-MLX-4bit,41619c55b00b41f7900c7d60c66eb0c790b63a6d,https://huggingface.co/Xylolabs/Hy4-MLX-4bit/tree/41619c55b00b41f7900c7d60c66eb0c790b63a6d,Exact sum of safetensors weight bytes at pinned Hub revision,Corrected full aggregate; no exclusions or batch-bug rows,yes
3
+ Xylolabs,MLX 4-bit,447576537740,0.19301077975582598,90.1256913021619,Chat evaluation text,512,0-38,39,256-510,9945,11.838745453335,tencent/Hy4 BF16 GGUF with native FP8 Engram; checked redo bases,c10c0de5b6f850703fb0a90e65d1c32b558b9167,12288,12288,llama.cpp; checked BIG_BASES2 redo completed 2026-11-21,MLX 0.32.4; mlx-lm 0.32.2; CUDA 13; bf16 activations; fp32 LM head; zero cache; sequential layer loading,1,f84b78403f3719ae91118fff277c8c17da876545c2be32f98734b062df00fb5b,3766cffb20a064f19e033e41bd9617fe9157d12cebaff439ca5055f1b96731d1,Xylolabs/Hy4-MLX-4bit,41619c55b00b41f7900c7d60c66eb0c790b63a6d,https://huggingface.co/Xylolabs/Hy4-MLX-4bit/tree/41619c55b00b41f7900c7d60c66eb0c790b63a6d,Exact sum of safetensors weight bytes at pinned Hub revision,Corrected full aggregate; no exclusions or batch-bug rows,no
4
+ Xylolabs,MLX 4-bit,447576537740,0.06618458081589519,92.36274509803921,Held-out evaluation text,512,0-39,40,256-510,10200,3.7460758857742245,tencent/Hy4 BF16 GGUF with native FP8 Engram; checked redo bases,c10c0de5b6f850703fb0a90e65d1c32b558b9167,12288,12288,llama.cpp; checked BIG_BASES2 redo completed 2026-11-21,MLX 0.32.4; mlx-lm 0.32.2; CUDA 13; bf16 activations; fp32 LM head; zero cache; sequential layer loading,1,4d3d0d85545c1738bf0b7e0c1d1cd74f09523c27d34d9e1a9c51e44f3de4c665,3766cffb20a064f19e033e41bd9617fe9157d12cebaff439ca5055f1b96731d1,Xylolabs/Hy4-MLX-4bit,41619c55b00b41f7900c7d60c66eb0c790b63a6d,https://huggingface.co/Xylolabs/Hy4-MLX-4bit/tree/41619c55b00b41f7900c7d60c66eb0c790b63a6d,Exact sum of safetensors weight bytes at pinned Hub revision,Corrected full aggregate; no exclusions or batch-bug rows,no
5
+ Xylolabs,MLX 8-bit,819628080300,,,WikiText-2 raw test,512,0-79,80,256-510,20400,,tencent/Hy4 BF16 GGUF with native FP8 Engram; checked redo bases,c10c0de5b6f850703fb0a90e65d1c32b558b9167,12288,12288,llama.cpp; checked BIG_BASES2 redo completed 2026-11-21,MLX 0.32.4; mlx-lm 0.32.2; CUDA 13; bf16 activations; fp32 LM head; zero cache; sequential layer loading,1,173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08,,Xylolabs/Hy4-MLX-8bit,cbce8a7f0cde8ca47731102c596a41208932c4de,https://huggingface.co/Xylolabs/Hy4-MLX-8bit/tree/cbce8a7f0cde8ca47731102c596a41208932c4de,Exact sum of safetensors weight bytes at pinned Hub revision,Pending; no completed corrected 8-bit measurement,yes
6
+ Xylolabs,MLX 8-bit,819628080300,,,Chat evaluation text,512,0-38,39,256-510,9945,,tencent/Hy4 BF16 GGUF with native FP8 Engram; checked redo bases,c10c0de5b6f850703fb0a90e65d1c32b558b9167,12288,12288,llama.cpp; checked BIG_BASES2 redo completed 2026-11-21,MLX 0.32.4; mlx-lm 0.32.2; CUDA 13; bf16 activations; fp32 LM head; zero cache; sequential layer loading,1,f84b78403f3719ae91118fff277c8c17da876545c2be32f98734b062df00fb5b,,Xylolabs/Hy4-MLX-8bit,cbce8a7f0cde8ca47731102c596a41208932c4de,https://huggingface.co/Xylolabs/Hy4-MLX-8bit/tree/cbce8a7f0cde8ca47731102c596a41208932c4de,Exact sum of safetensors weight bytes at pinned Hub revision,Pending; no completed corrected 8-bit measurement,no
7
+ Xylolabs,MLX 8-bit,819628080300,,,Held-out evaluation text,512,0-39,40,256-510,10200,,tencent/Hy4 BF16 GGUF with native FP8 Engram; checked redo bases,c10c0de5b6f850703fb0a90e65d1c32b558b9167,12288,12288,llama.cpp; checked BIG_BASES2 redo completed 2026-11-21,MLX 0.32.4; mlx-lm 0.32.2; CUDA 13; bf16 activations; fp32 LM head; zero cache; sequential layer loading,1,4d3d0d85545c1738bf0b7e0c1d1cd74f09523c27d34d9e1a9c51e44f3de4c665,,Xylolabs/Hy4-MLX-8bit,cbce8a7f0cde8ca47731102c596a41208932c4de,https://huggingface.co/Xylolabs/Hy4-MLX-8bit/tree/cbce8a7f0cde8ca47731102c596a41208932c4de,Exact sum of safetensors weight bytes at pinned Hub revision,Pending; no completed corrected 8-bit measurement,no
accuracy-kl-dark.png ADDED
accuracy-kl.png ADDED
accuracy-top1-dark.png ADDED
accuracy-top1.png ADDED