dahara1 commited on
Commit
391c292
1 Parent(s): 13acbee

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +10 -4
README.md CHANGED
@@ -33,13 +33,19 @@ Measurements using Japanese ja-wiki.test.raw data showed that L/fp16 quants cont
33
 
34
  ## 用語集 Terminology
35
 
 
 
 
 
 
 
36
  wiki.test.raw score
37
- Perplexity Score measured using [wiki.test.raw](https://huggingface.co/datasets/Salesforce/wikitext/viewer/wikitext-103-raw-v1/test) published by Salesforce and the [llama-perplexity](https://github.com/ggerganov/llama.cpp/blob/master/examples/perplexity/README.md) command with -c 512 setting.
38
- Salesforceが公開してくれている[wiki.test.raw](https://huggingface.co/datasets/Salesforce/wikitext/viewer/wikitext-103-raw-v1/test)と[llama-perplexity](https://github.com/ggerganov/llama.cpp/blob/master/examples/perplexity/README.md)コマンドの -c 512設定で計測した数値
39
 
40
  ja-wiki.test.raw perplexity score
41
- Perplexity score measured in the same way using Japanese data ([ja-wiki.test.raw](https://huggingface.co/dahara1/imatrix-jpn-test/blob/main/ja-wiki.test.raw)) with the same file size as wiki.test.raw, extracted from a Japanese wiki.
42
- 日本語のwikiから抜き出した文章でwiki.test.rawと同じファイルサイズにした日本語データ([ja-wiki.test.raw](https://huggingface.co/dahara1/imatrix-jpn-test/blob/main/ja-wiki.test.raw))で測定したperplexity score.
43
 
44
  Bartowski model
45
  [Bartowski](https://huggingface.co/bartowski) is an individual who has quantized many models and contributed to the community. He created an imatrix from the English-only data [calibration_datav3.txt](https://gist.github.com/bartowski1182/eb213dccb3571f863da82e99418f81e8) and used it to quantize his model.
 
33
 
34
  ## 用語集 Terminology
35
 
36
+ Importance Matrix (imatrix)
37
+ An "imatrix" is a data structure used to optimize the quantization of a model. You create one by passing text to the [llama-imatrix](https://github.com/ggerganov/llama.cpp/tree/master/examples/imatrix) command.
38
+
39
+ 重要行列(imatrix)
40
+ 「imatrix」は、モデルの量子化を最適化するために使用されるデータ構造です。[llama-imatrix](https://github.com/ggerganov/llama.cpp/tree/master/examples/imatrix)コマンドにテキストを与えて作成します
41
+
42
  wiki.test.raw score
43
+ Perplexity Score measured using [wiki.test.raw](https://huggingface.co/datasets/Salesforce/wikitext/viewer/wikitext-103-raw-v1/test) published by Salesforce and the [llama-perplexity](https://github.com/ggerganov/llama.cpp/blob/master/examples/perplexity/README.md) command with -c 512 setting. Lower values are better.
44
+ Salesforceが公開してくれている[wiki.test.raw](https://huggingface.co/datasets/Salesforce/wikitext/viewer/wikitext-103-raw-v1/test)と[llama-perplexity](https://github.com/ggerganov/llama.cpp/blob/master/examples/perplexity/README.md)コマンドの -c 512設定で計測した数値。値が低いほど優れています。
45
 
46
  ja-wiki.test.raw perplexity score
47
+ Perplexity score measured in the same way using Japanese data ([ja-wiki.test.raw](https://huggingface.co/dahara1/imatrix-jpn-test/blob/main/ja-wiki.test.raw)) with the same file size as wiki.test.raw, extracted from a Japanese wiki. Lower values are better.
48
+ 日本語のwikiから抜き出した文章でwiki.test.rawと同じファイルサイズにした日本語データ([ja-wiki.test.raw](https://huggingface.co/dahara1/imatrix-jpn-test/blob/main/ja-wiki.test.raw))で測定したperplexity score. 値が低いほど優れています。
49
 
50
  Bartowski model
51
  [Bartowski](https://huggingface.co/bartowski) is an individual who has quantized many models and contributed to the community. He created an imatrix from the English-only data [calibration_datav3.txt](https://gist.github.com/bartowski1182/eb213dccb3571f863da82e99418f81e8) and used it to quantize his model.