File size: 1,719 Bytes
bffc5ed 4fc7faa bffc5ed 4fc7faa bffc5ed cb13c9b 4fc7faa bffc5ed 4fc7faa bffc5ed 5a26c8b 4fc7faa bffc5ed 4fc7faa bffc5ed 4fc7faa |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 |
---
license: other
language:
- ja
base_model:
- deepseek-ai/DeepSeek-V3
---
# DeepSeek-V3-slice-jp64
## 実験モデルです
本モデルは [DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3) をベースに、日本語の例文を元に頻出する MoE (Mixture of Experts) の各レイヤーごとのexpertsを厳選して再構成したモデルです。
元のモデルでは 256 のexpertsを搭載していますが、日本語出力における安定性とパフォーマンスのバランスを重視し、各層で頻出する 64 のexpertsを使用するように調整しています。
### 例文出力時の各layerごとのexpertsの頻出分布
![](layer_topk_idx_distribution_bubble.png)
---
## ライセンス
ご使用前にライセンスファイルをご確認ください。
[DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3) こちらのライセンスをそのまま使用しています。
## 特徴
- MoEモデルのexpertsから、日本語の例文出力をして各layerごとに頻出する64のexpertをして組み直したモデルです。
- 16ではまともに動かず、32では安定しなかったため64expertsにしています。
- scripts/layer_topk_idx_distribution.json
- 各layerごとに頻出順に128のexpertのrankが記録されています。
- scripts/deepseek_slice.py
- 元モデル(bf16)から、64のexpertを使用したモデル(bf16)を作成します。
- scripts/model_test.py
- モデル実行用テスト用のスクリプトです。コメントアウトされている例文を元に頻出するexpertを計測しています
---
## 使い方
`scripts/model_test.py`に実行コードあります |