Update README.md
Browse files
README.md
CHANGED
@@ -23,20 +23,31 @@ This page confirms the effectiveness of multilingual imatrix.
|
|
23 |
|
24 |
|
25 |
![wiki.test.raw_perplexity_score.png](wiki.test.raw_perplexity_score.png)
|
26 |
-
Measurements using English wiki.test.raw
|
27 |
-
英語のwiki.test.rawを使った計測ではimatrix
|
28 |
|
29 |
![ja-wiki.test.raw_perplexity_score.png](ja-wiki.test.raw_perplexity_score.png)
|
30 |
|
31 |
-
Measurements using Japanese ja-wiki.test.raw data
|
32 |
-
日本語のja-wiki.test.rawデータを使った計測ではL/fp16
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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
|
|
|
23 |
|
24 |
|
25 |
![wiki.test.raw_perplexity_score.png](wiki.test.raw_perplexity_score.png)
|
26 |
+
Measurements using English wiki.test.raw suggest that imatrix improves perplexity scores.
|
27 |
+
英語のwiki.test.rawを使った計測ではimatrixがperplexityスコアを向上させる事が示唆された
|
28 |
|
29 |
![ja-wiki.test.raw_perplexity_score.png](ja-wiki.test.raw_perplexity_score.png)
|
30 |
|
31 |
+
Measurements using Japanese ja-wiki.test.raw data suggest that L/fp16 quants improve scores.
|
32 |
+
日本語のja-wiki.test.rawデータを使った計測ではL/fp16クォンツがスコアを向上させる事が示唆された
|
33 |
+
|
34 |
+
| Model | wiki.test.raw Perplexity | ja-wiki.test.raw Perplexity |
|
35 |
+
|----------------------|--------------------------|-----------------------------|
|
36 |
+
| bartowski M | 8.8140 | 17.2091 |
|
37 |
+
| bartowski L | 8.8137 | 17.1035 |
|
38 |
+
| bartowski fp16 | 8.8146 | 17.0946 |
|
39 |
+
| imatrix-jpn-test M | 8.8231 | 17.2069 |
|
40 |
+
| imatrix-jpn-test L | 8.8193 | 17.0931 |
|
41 |
+
| imatrix-jpn-test fp16| 8.8198 | 17.0877 |
|
42 |
+
| no imatrix M | 8.8885 | 17.3948 |
|
43 |
+
| no imatrix L | 8.8938 | 17.2974 |
|
44 |
+
| no imatrix fp16 | 8.8887 | 17.2740 |
|
45 |
+
|
46 |
|
47 |
## 用語集 Terminology
|
48 |
|
49 |
Importance Matrix (imatrix)
|
50 |
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.
|
|
|
|
|
51 |
「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
|
52 |
|
53 |
wiki.test.raw score
|