Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,52 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
+
language:
|
4 |
+
- ja
|
5 |
+
pipeline_tag: text-generation
|
6 |
+
tags:
|
7 |
+
- Mistral
|
8 |
---
|
9 |
+
|
10 |
+
# Japanese-WizardLM2-ChatV-7B-GGUF
|
11 |
+
GGUF conversion of "Japanese-WizardLM2-ChatV-7B"
|
12 |
+
|
13 |
+
This model, Japanese-WizardLM2-ChatV-7B, is based on "chatntq-ja-7b-v1.0 ", and was created by subtracting "Mistral-7B-v0.1" from "WizardLM-2-7b" ChatVector was added by a factor of 1.0.
|
14 |
+
We aimed to add the high performance of WizardLM-2 to the Japanese language capability of ChatNTQ.
|
15 |
+
|
16 |
+
このモデル、Japanese-WizardLM2-ChatV-7Bは、”chatntq-ja-7b-v1.0”をベースに、"WizardLM-2-7b"から"Mistral-7B-v0.1"を差し引いて作ったChatVectorを1.0倍で足しました。
|
17 |
+
ChatNTQの日本語能力にWizardLM-2の性能の高さが加わる事を狙いました。
|
18 |
+
|
19 |
+
### Performance
|
20 |
+
|
21 |
+
<table>
|
22 |
+
<tr>
|
23 |
+
<th>Model<br>(Q8_0 quant)</th>
|
24 |
+
<th><a href="https://huggingface.co/CohereForAI/c4ai-command-r-plus">c4ai-command-r-plus(Cohere API)</a></th>
|
25 |
+
<th><a href="https://huggingface.co/TFMC/Japanese-Starling-ChatV-7B-GGUF">JA-Starling-ChatV-7B-GGUF(fp16)</th>
|
26 |
+
<th>JA-WizardLM2-ChatV-7B-GGUF (This model)(fp16)</th>
|
27 |
+
<th><a href="https://huggingface.co/TFMC/ChatNTQ-JA-7b-v1.0-GGUF">ChatNTQ-JA-7b-v1.0-GGUF(Q8_0)</a></th>
|
28 |
+
</tr>
|
29 |
+
<tr>
|
30 |
+
<td>Parameters</td>
|
31 |
+
<td>104B</td>
|
32 |
+
<td>7B(Mistral)</td>
|
33 |
+
<td>7B(Mistral)</td>
|
34 |
+
<td>7B(Mistral)</td>
|
35 |
+
</tr>
|
36 |
+
<tr>
|
37 |
+
<td>ELYZAtasks100<br>average score</td>
|
38 |
+
<td>4.04</td>
|
39 |
+
<td>3.77</td>
|
40 |
+
<td>3.40</td>
|
41 |
+
<td>2.74</td>
|
42 |
+
</tr>
|
43 |
+
</table>
|
44 |
+
|
45 |
+
This score was evaluated by Cohere API command-r-plus using the "<a href="https://huggingface.co/datasets/elyza/ELYZA-tasks-100">ELYZA-tasks-100</a>", a Japanese model with instruction-tuning.
|
46 |
+
|
47 |
+
このスコアはinstruction-tuningを行った日本語モデルのベンチマーク「ELYZA-tasks-100」を使い、Cohere APIのcommand-r-plusにより評価させたものです。
|
48 |
+
|
49 |
+
### Prompt Template
|
50 |
+
|
51 |
+
- Llama-2-Chat
|
52 |
+
<pre><code>[INST] <<SYS>>\nあなたは役に立つアシスタントです。\n<</SYS>>\n\n{prompt} [/INST]</code></pre>
|