Dracones commited on
Commit
650cf39
1 Parent(s): 57cc7ae

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +56 -0
README.md CHANGED
@@ -19,6 +19,62 @@ These quants were made with exllamav2 version 0.0.18. Quants made on this versio
19
 
20
  If you have problems loading these models, please update Text Generation WebUI to the latest version.
21
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
22
 
23
 
24
  ## Quant Details
 
19
 
20
  If you have problems loading these models, please update Text Generation WebUI to the latest version.
21
 
22
+ ## Perplexity Scoring
23
+
24
+ Below are the perplexity scores for the EXL2 models. A lower score is better.
25
+
26
+ | Quant Level | Perplexity Score |
27
+ |-------------|------------------|
28
+ | 7.0 | 4.5859 |
29
+ | 6.0 | 4.6252 |
30
+ | 5.5 | 4.6493 |
31
+ | 5.0 | 4.6937 |
32
+ | 4.5 | 4.8029 |
33
+ | 4.0 | 4.9372 |
34
+ | 3.5 | 5.1336 |
35
+ | 3.25 | 5.3636 |
36
+ | 3.0 | 5.5468 |
37
+ | 2.75 | 5.8255 |
38
+ | 2.5 | 6.3362 |
39
+ | 2.25 | 7.7763 |
40
+
41
+
42
+ ### Perplexity Script
43
+
44
+ This was the script used for perplexity testing.
45
+
46
+ ```bash
47
+ #!/bin/bash
48
+
49
+ # Activate the conda environment
50
+ source ~/miniconda3/etc/profile.d/conda.sh
51
+ conda activate exllamav2
52
+
53
+ DATA_SET=/root/wikitext/wikitext-2-v1.parquet
54
+
55
+ # Set the model name and bit size
56
+ MODEL_NAME="WizardLM-2-8x22B"
57
+ BIT_PRECISIONS=(6.0 5.5 5.0 4.5 4.0 3.5 3.25 3.0 2.75 2.5 2.25)
58
+
59
+ # Print the markdown table header
60
+ echo "| Quant Level | Perplexity Score |"
61
+ echo "|-------------|------------------|"
62
+
63
+ for BIT_PRECISION in "${BIT_PRECISIONS[@]}"
64
+ do
65
+ LOCAL_FOLDER="/root/models/${MODEL_NAME}_exl2_${BIT_PRECISION}bpw"
66
+ REMOTE_FOLDER="Dracones/${MODEL_NAME}_exl2_${BIT_PRECISION}bpw"
67
+
68
+ if [ ! -d "$LOCAL_FOLDER" ]; then
69
+ huggingface-cli download --local-dir-use-symlinks=False --local-dir "${LOCAL_FOLDER}" "${REMOTE_FOLDER}" >> /root/download.log 2>&1
70
+ fi
71
+
72
+ output=$(python test_inference.py -m "$LOCAL_FOLDER" -gs 40,40,40,40 -ed "$DATA_SET")
73
+ score=$(echo "$output" | grep -oP 'Evaluation perplexity: \K[\d.]+')
74
+ echo "| $BIT_PRECISION | $score |"
75
+ # rm -rf "${LOCAL_FOLDER}"
76
+ done
77
+ ```
78
 
79
 
80
  ## Quant Details