Update README.md
Browse files
README.md
CHANGED
@@ -25,4 +25,98 @@ datasets:
|
|
25 |
|
26 |
This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
|
27 |
|
28 |
-
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
|
26 |
This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
|
27 |
|
28 |
+
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
|
29 |
+
|
30 |
+
|
31 |
+
## Usage
|
32 |
+
Execute following code in Google Colab
|
33 |
+
|
34 |
+
```python
|
35 |
+
# 必要なライブラリをインストール
|
36 |
+
!pip install unsloth
|
37 |
+
!pip uninstall unsloth -y && pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
|
38 |
+
!pip install -U torch
|
39 |
+
!pip install -U peft
|
40 |
+
|
41 |
+
# 必要なライブラリを読み込み
|
42 |
+
from unsloth import FastLanguageModel
|
43 |
+
from peft import PeftModel
|
44 |
+
import torch
|
45 |
+
import json
|
46 |
+
from tqdm import tqdm
|
47 |
+
import re
|
48 |
+
|
49 |
+
# ベースとなるモデルと学習したLoRAのアダプタ(Hugging FaceのIDを指定)。
|
50 |
+
model_id = "llm-jp/llm-jp-3-13b"
|
51 |
+
adapter_id = "Ken5615/llm-jp-3-13b-ft-bigest-latest"
|
52 |
+
|
53 |
+
from google.colab import userdata
|
54 |
+
HF_TOKEN=userdata.get('HF_TOKEN')
|
55 |
+
|
56 |
+
# unslothのFastLanguageModelで元のモデルをロード。
|
57 |
+
dtype = None # Noneにしておけば自動で設定
|
58 |
+
load_in_4bit = True # 今回は13Bモデルを扱うためTrue
|
59 |
+
|
60 |
+
model, tokenizer = FastLanguageModel.from_pretrained(
|
61 |
+
model_name=model_id,
|
62 |
+
dtype=dtype,
|
63 |
+
load_in_4bit=load_in_4bit,
|
64 |
+
trust_remote_code=True,
|
65 |
+
)
|
66 |
+
|
67 |
+
# 元のモデルにLoRAのアダプタを統合。
|
68 |
+
model = PeftModel.from_pretrained(model, adapter_id, token = HF_TOKEN)
|
69 |
+
|
70 |
+
# タスクとなるデータの読み込み。
|
71 |
+
# 事前にデータをアップロードしてください。
|
72 |
+
datasets = []
|
73 |
+
with open("./elyza-tasks-100-TV_0.jsonl", "r") as f:
|
74 |
+
item = ""
|
75 |
+
for line in f:
|
76 |
+
line = line.strip()
|
77 |
+
item += line
|
78 |
+
if item.endswith("}"):
|
79 |
+
datasets.append(json.loads(item))
|
80 |
+
item = ""
|
81 |
+
|
82 |
+
# モデルを用いてタスクの推論。
|
83 |
+
|
84 |
+
# 推論するためにモデルのモードを変更
|
85 |
+
FastLanguageModel.for_inference(model)
|
86 |
+
|
87 |
+
results = []
|
88 |
+
for dt in tqdm(datasets):
|
89 |
+
input = dt["input"]
|
90 |
+
|
91 |
+
prompt = f"""### 以下の入力の内容に以下の条件を守って回答してください。\n### 条件\n・step-by-stepで考えて回答してください。\n・入力に示された回答方法を守って回答してください。\n### 入力\n{input}\n### 回答\n"""
|
92 |
+
|
93 |
+
|
94 |
+
inputs = tokenizer([prompt], return_tensors = "pt").to(model.device)
|
95 |
+
|
96 |
+
outputs = model.generate(**inputs, max_new_tokens = 1024, use_cache = True, do_sample=False, repetition_penalty=1.2)
|
97 |
+
prediction = tokenizer.decode(outputs[0], skip_special_tokens=True).split('\n### 回答')[-1]
|
98 |
+
prediction = re.sub(r"[*#]", "", prediction)
|
99 |
+
|
100 |
+
results.append({"task_id": dt["task_id"], "input": input, "output": prediction})
|
101 |
+
|
102 |
+
# 結果をjsonlで保存。
|
103 |
+
json_file_id = re.sub(".*/", "", adapter_id)
|
104 |
+
with open(f"/content/{json_file_id}_output.jsonl", 'w', encoding='utf-8') as f:
|
105 |
+
for result in results:
|
106 |
+
json.dump(result, f, ensure_ascii=False)
|
107 |
+
f.write('\n')
|
108 |
+
```
|
109 |
+
|
110 |
+
|
111 |
+
## Datasets
|
112 |
+
|
113 |
+
### Instruction tuning
|
114 |
+
|
115 |
+
The models have been fine-tuned on the following datasets.
|
116 |
+
|
117 |
+
| Language | Dataset | description |
|
118 |
+
|:---|:---|:---|
|
119 |
+
|Japanese|[ichikara-instruction-004-002](https://liat-aip.sakura.ne.jp/wp/llm%e3%81%ae%e3%81%9f%e3%82%81%e3%81%ae%e6%97%a5%e6%9c%ac%e8%aa%9e%e3%82%a4%e3%83%b3%e3%82%b9%e3%83%88%e3%83%a9%e3%82%af%e3%82%b7%e3%83%a7%e3%83%b3%e3%83%87%e3%83%bc%e3%82%bf%e4%bd%9c%e6%88%90/llm%e3%81%ae%e3%81%9f%e3%82%81%e3%81%ae%e6%97%a5%e6%9c%ac%e8%aa%9e%e3%82%a4%e3%83%b3%e3%82%b9%e3%83%88%e3%83%a9%e3%82%af%e3%82%b7%e3%83%a7%e3%83%b3%e3%83%87%e3%83%bc%e3%82%bf-%e5%85%ac%e9%96%8b/)| A manually constructed instruction dataset |
|
120 |
+
| |[llm-jp/databricks-dolly-15k-ja](https://huggingface.co/datasets/llm-jp/databricks-dolly-15k-ja)| A manually constructed instruction dataset created by llm-jp|
|
121 |
+
| |[elyza/ELYZA-tasks-100](https://huggingface.co/datasets/elyza/ELYZA-tasks-100)| A manually constructed instruction dataset created by elyza|
|
122 |
+
| | Synthesized data from [elyza/ELYZA-tasks-100](https://huggingface.co/datasets/elyza/ELYZA-tasks-100)| Synthesize data from [Elyza-tasks-100](https://huggingface.co/datasets/elyza/ELYZA-tasks-100) by using LLM(Tanuki-8x8B) |
|