Hugging Face's logo

ngquocvinh
/
Humanizer-12B-GGUF

Text Generation
GGUF
English
Chinese
llama.cpp
humanizer
quantized
imatrix
conversational
Model card Files
xet
 Community
ngquocvinh commited on
Commit
34492ac
·
verified ·
1 Parent(s): 7404c8d

docs: simplify artifact measurement table

Browse files
Files changed (3) hide show
  1. README.md +25 -25
  2. reproducibility/build-measurements.tsv +19 -19
  3. reproducibility/manifest.md +1 -1
README.md CHANGED
@@ -34,30 +34,30 @@ These are direct GGUF quantizations of the upstream BF16 weights. No additional
34
 
35
  ## Artifact measurements
36
 
37
- Sizes are measured from the local files in decimal GB (`bytes / 1,000,000,000`). Quantization time is wall time per artifact using 8 CPU threads. Peak build RSS + cache is the observed cgroup estimate (process RSS plus charged file cache), in GiB. The BF16 row links to the upstream reference and was not quantized locally.
38
-
39
- | File | Size (GB) | Quantization time (s) | Peak build RSS + cache (GiB) | Runtime smoke |
40
- |---|---:|---:|---:|---|
41
- | [humanizer-12b-bf16.gguf (upstream reference)](https://huggingface.co/jialinyyzz/humanizer/resolve/766072b26fbea83778a0ddc2dd577498098670bd/humanizer-12b-bf16.gguf) | 23.832050 | N/A | N/A | PASS (EN + ZH); upstream BF16 reference |
42
- | `humanizer-12b-Q8_0.gguf` | 12.669631 | 45.045 | 12.156696 | PASS (EN + ZH) |
43
- | `humanizer-12b-Q6_K.gguf` | 9.786006 | 95.041 | 13.775974 | PASS (EN + ZH) |
44
- | `humanizer-12b-Q5_K_M.gguf` | 8.547252 | 210.081 | 12.997784 | PASS (EN + ZH) |
45
- | `humanizer-12b-Q4_K_M.gguf` | 7.381367 | 205.085 | 11.978336 | PASS (EN + ZH) |
46
- | `humanizer-12b-Q3_K_M.gguf` | 6.087072 | 145.061 | 10.795521 | PASS (EN + ZH) |
47
- | `humanizer-12b-Q2_K.gguf` | 4.830132 | 220.080 | 9.681194 | PASS (EN + ZH); low-bit experimental |
48
- | `humanizer-12b-IQ2_XS.gguf` | 3.887827 | 1040.381 | 8.917431 | PASS (EN + ZH); low-bit experimental |
49
- | `humanizer-12b-IQ1_M.gguf` | 3.202833 | 315.110 | 14.173813 | PASS (EN + ZH); low-bit experimental |
50
- | `humanizer-12b-Q1_0.gguf` | 2.377448 | 35.017 | 13.498875 | PASS (EN + ZH); low-bit experimental |
51
- | `humanizer-12b-Q5_K_S.gguf` | 8.338356 | 230.102 | 18.925488 | PASS (EN + ZH) |
52
- | `humanizer-12b-Q4_K_S.gguf` | 7.024032 | 225.083 | 17.732090 | PASS (EN + ZH) |
53
- | `humanizer-12b-Q3_K_S.gguf` | 5.528214 | 95.044 | 16.260834 | PASS (EN + ZH) |
54
- | `humanizer-12b-Q3_K_L.gguf` | 6.566304 | 145.058 | 17.284103 | PASS (EN + ZH) |
55
- | `humanizer-12b-IQ4_XS.gguf` | 6.635240 | 420.165 | 17.493828 | PASS (EN + ZH) |
56
- | `humanizer-12b-IQ3_XXS.gguf` | 4.849179 | 700.222 | 15.862881 | PASS (EN + ZH) |
57
- | `humanizer-12b-IQ2_XXS.gguf` | 3.568093 | 570.202 | 14.703056 | PASS (EN + ZH); low-bit experimental |
58
- | `humanizer-12b-TQ2_0.gguf` | 3.654785 | 30.015 | 14.486164 | PASS (EN + ZH); low-bit experimental |
59
-
60
- The bilingual CPU smoke test loads each listed file through llama.cpp and confirms a non-empty generation for one English and one Chinese draft. These build and smoke measurements do not measure inference speed or output fidelity. No hold-out fidelity scores are published.
61
 
62
  ## Quick start
63
 
@@ -98,7 +98,7 @@ For short English drafts, upstream notes that a Chinese reply can occasionally o
98
  ## Reproducibility and validation
99
 
100
  - [Artifact measurements](reproducibility/artifact-measurements.tsv) report exact bytes and decimal GB.
101
- - [Build measurements](reproducibility/build-measurements.tsv) report per-format quantization time, peak RSS-plus-file-cache, and bilingual CPU smoke status.
102
  - [Reproducibility manifest](reproducibility/manifest.md) records the source revision, BF16 and prompt hashes, calibration split, imatrix, quantization commands, and runtime checks.
103
  - [SHA256 checksums](SHA256SUMS.txt) cover the published GGUF files, prompt format, and license.
104
 
 
34
 
35
  ## Artifact measurements
36
 
37
+ Sizes are measured from the local files in decimal GB (`bytes / 1,000,000,000`). Peak build RSS + cache is the observed cgroup estimate (process RSS plus charged file cache), in GiB. The BF16 row links to the upstream reference and was not quantized locally.
38
+
39
+ | File | Size (GB) | Peak build RSS + cache (GiB) |
40
+ |---|---:|---:|
41
+ | [humanizer-12b-bf16.gguf (upstream reference)](https://huggingface.co/jialinyyzz/humanizer/resolve/766072b26fbea83778a0ddc2dd577498098670bd/humanizer-12b-bf16.gguf) | 23.832050 | N/A |
42
+ | `humanizer-12b-Q8_0.gguf` | 12.669631 | 12.156696 |
43
+ | `humanizer-12b-Q6_K.gguf` | 9.786006 | 13.775974 |
44
+ | `humanizer-12b-Q5_K_M.gguf` | 8.547252 | 12.997784 |
45
+ | `humanizer-12b-Q4_K_M.gguf` | 7.381367 | 11.978336 |
46
+ | `humanizer-12b-Q3_K_M.gguf` | 6.087072 | 10.795521 |
47
+ | `humanizer-12b-Q2_K.gguf` | 4.830132 | 9.681194 |
48
+ | `humanizer-12b-IQ2_XS.gguf` | 3.887827 | 8.917431 |
49
+ | `humanizer-12b-IQ1_M.gguf` | 3.202833 | 14.173813 |
50
+ | `humanizer-12b-Q1_0.gguf` | 2.377448 | 13.498875 |
51
+ | `humanizer-12b-Q5_K_S.gguf` | 8.338356 | 18.925488 |
52
+ | `humanizer-12b-Q4_K_S.gguf` | 7.024032 | 17.732090 |
53
+ | `humanizer-12b-Q3_K_S.gguf` | 5.528214 | 16.260834 |
54
+ | `humanizer-12b-Q3_K_L.gguf` | 6.566304 | 17.284103 |
55
+ | `humanizer-12b-IQ4_XS.gguf` | 6.635240 | 17.493828 |
56
+ | `humanizer-12b-IQ3_XXS.gguf` | 4.849179 | 15.862881 |
57
+ | `humanizer-12b-IQ2_XXS.gguf` | 3.568093 | 14.703056 |
58
+ | `humanizer-12b-TQ2_0.gguf` | 3.654785 | 14.486164 |
59
+
60
+ Peak build memory and file size describe artifact/build footprint, not inference speed or output fidelity. Each listed file passed a bilingual CPU smoke test with a non-empty generation for one English and one Chinese draft. No hold-out fidelity scores are published.
61
 
62
  ## Quick start
63
 
 
98
  ## Reproducibility and validation
99
 
100
  - [Artifact measurements](reproducibility/artifact-measurements.tsv) report exact bytes and decimal GB.
101
+ - [Build measurements](reproducibility/build-measurements.tsv) report per-format peak RSS-plus-file-cache.
102
  - [Reproducibility manifest](reproducibility/manifest.md) records the source revision, BF16 and prompt hashes, calibration split, imatrix, quantization commands, and runtime checks.
103
  - [SHA256 checksums](SHA256SUMS.txt) cover the published GGUF files, prompt format, and license.
104
 
reproducibility/build-measurements.tsv CHANGED
@@ -1,19 +1,19 @@
1
- file size_bytes size_gb quantization_seconds peak_rss_plus_file_cache_gib cpu_smoke
2
- humanizer-12b-bf16.gguf (upstream reference) 23832049568 23.832050 N/A N/A PASS_EN_ZH_BF16_REFERENCE
3
- humanizer-12b-Q8_0.gguf 12669630688 12.669631 45.045 12.156696 PASS_EN_ZH
4
- humanizer-12b-Q6_K.gguf 9786005728 9.786006 95.041 13.775974 PASS_EN_ZH
5
- humanizer-12b-Q5_K_M.gguf 8547252448 8.547252 210.081 12.997784 PASS_EN_ZH
6
- humanizer-12b-Q4_K_M.gguf 7381367008 7.381367 205.085 11.978336 PASS_EN_ZH
7
- humanizer-12b-Q3_K_M.gguf 6087071968 6.087072 145.061 10.795521 PASS_EN_ZH
8
- humanizer-12b-Q2_K.gguf 4830132448 4.830132 220.080 9.681194 PASS_EN_ZH
9
- humanizer-12b-IQ2_XS.gguf 3887827168 3.887827 1040.381 8.917431 PASS_EN_ZH
10
- humanizer-12b-IQ1_M.gguf 3202832608 3.202833 315.110 14.173813 PASS_EN_ZH
11
- humanizer-12b-Q1_0.gguf 2377447648 2.377448 35.017 13.498875 PASS_EN_ZH
12
- humanizer-12b-Q5_K_S.gguf 8338356448 8.338356 230.102 18.925488 PASS_EN_ZH
13
- humanizer-12b-Q4_K_S.gguf 7024031968 7.024032 225.083 17.732090 PASS_EN_ZH
14
- humanizer-12b-Q3_K_S.gguf 5528213728 5.528214 95.044 16.260834 PASS_EN_ZH
15
- humanizer-12b-Q3_K_L.gguf 6566303968 6.566304 145.058 17.284103 PASS_EN_ZH
16
- humanizer-12b-IQ4_XS.gguf 6635239648 6.635240 420.165 17.493828 PASS_EN_ZH
17
- humanizer-12b-IQ3_XXS.gguf 4849178848 4.849179 700.222 15.862881 PASS_EN_ZH
18
- humanizer-12b-IQ2_XXS.gguf 3568093408 3.568093 570.202 14.703056 PASS_EN_ZH
19
- humanizer-12b-TQ2_0.gguf 3654785248 3.654785 30.015 14.486164 PASS_EN_ZH
 
1
+ file size_bytes size_gb peak_rss_plus_file_cache_gib
2
+ humanizer-12b-bf16.gguf (upstream reference) 23832049568 23.832050 N/A
3
+ humanizer-12b-Q8_0.gguf 12669630688 12.669631 12.156696
4
+ humanizer-12b-Q6_K.gguf 9786005728 9.786006 13.775974
5
+ humanizer-12b-Q5_K_M.gguf 8547252448 8.547252 12.997784
6
+ humanizer-12b-Q4_K_M.gguf 7381367008 7.381367 11.978336
7
+ humanizer-12b-Q3_K_M.gguf 6087071968 6.087072 10.795521
8
+ humanizer-12b-Q2_K.gguf 4830132448 4.830132 9.681194
9
+ humanizer-12b-IQ2_XS.gguf 3887827168 3.887827 8.917431
10
+ humanizer-12b-IQ1_M.gguf 3202832608 3.202833 14.173813
11
+ humanizer-12b-Q1_0.gguf 2377447648 2.377448 13.498875
12
+ humanizer-12b-Q5_K_S.gguf 8338356448 8.338356 18.925488
13
+ humanizer-12b-Q4_K_S.gguf 7024031968 7.024032 17.732090
14
+ humanizer-12b-Q3_K_S.gguf 5528213728 5.528214 16.260834
15
+ humanizer-12b-Q3_K_L.gguf 6566303968 6.566304 17.284103
16
+ humanizer-12b-IQ4_XS.gguf 6635239648 6.635240 17.493828
17
+ humanizer-12b-IQ3_XXS.gguf 4849178848 4.849179 15.862881
18
+ humanizer-12b-IQ2_XXS.gguf 3568093408 3.568093 14.703056
19
+ humanizer-12b-TQ2_0.gguf 3654785248 3.654785 14.486164
reproducibility/manifest.md CHANGED
@@ -21,7 +21,7 @@
21
 
22
  ## Quantization and runtime validation
23
 
24
- Each published GGUF is quantized directly from the locked BF16 GGUF using the model-specific imatrix. Each artifact is smoke-tested as a text-completion model via llama.cpp on CPU. The test preserves the prompt format and checks non-empty English and Chinese generations. The public measurements report file sizes, quantization wall time, peak build RSS-plus-file-cache, and runtime smoke status; they do not measure inference speed or held-out fidelity.
25
 
26
  ### BF16 upstream reference
27
 
 
21
 
22
  ## Quantization and runtime validation
23
 
24
+ Each published GGUF is quantized directly from the locked BF16 GGUF using the model-specific imatrix. Each artifact is smoke-tested as a text-completion model via llama.cpp on CPU. The test preserves the prompt format and checks non-empty English and Chinese generations. The public measurements report file sizes and peak build RSS-plus-file-cache; they do not measure inference speed or held-out fidelity.
25
 
26
  ### BF16 upstream reference
27