|
|
--- |
|
|
library_name: transformers |
|
|
license: mit |
|
|
datasets: |
|
|
- HuggingFaceH4/Bespoke-Stratos-17k |
|
|
- AI-MO/NuminaMath-TIR |
|
|
language: |
|
|
- en |
|
|
- ja |
|
|
base_model: |
|
|
- microsoft/phi-4 |
|
|
--- |
|
|
|
|
|
 |
|
|
|
|
|
# AXCXEPT/phi-4-open-R1-Distill-EZOv1 |
|
|
|
|
|
<!-- Provide a quick summary of what the model is/does. --> |
|
|
|
|
|
## Model Details |
|
|
This model is a Reasoner version of the phi-4 model by employing open-r1, which mimics the Distill methodology of Deepseek-R1. In particular, since it is specialized for Japanese, answers will be given in Japanese in principle. |
|
|
It can be made more flexible by interweaving English. |
|
|
|
|
|
Deepseek-R1のDistill手法を模倣した、open-r1を採用して、phi-4モデルを Reasonerにしたモデルです。特に日本語に特化させているため、原則日本語で回答します。 |
|
|
英語を織り交ぜることでより柔軟にすることが可能です。 |
|
|
|
|
|
## Example Output |
|
|
``` |
|
|
<thinking> |
|
|
まず、問題文を整理します。父親の現在の年齢は息子の3倍であり、6年前の父親の年齢は息子の5倍でした。これを数式で表現すると： |
|
|
|
|
|
1. 現在の年齢関係：父親の年齢 = 3 × 息子の年齢 → F = 3S |
|
|
2. 6年前の年齢関係：父親の年齢 - 6 = 5 × (息子の年齢 - 6) → F - 6 = 5(S - 6) |
|
|
|
|
|
これらの2つの方程式を同時に解きます。まず第1式からF = 3Sを代入すると： |
|
|
|
|
|
3S - 6 = 5(S - 6) |
|
|
3S - 6 = 5S - 30 |
|
|
30 - 6 = 5S - 3S |
|
|
24 = 2S |
|
|
S = 12 |
|
|
|
|
|
息子の年齢が12歳であることが分かります。これを第1式に代入すると： |
|
|
|
|
|
F = 3 × 12 = 36 |
|
|
|
|
|
父親の年齢は36歳です。最後に条件を確認します： |
|
|
- 現在の年齢関係：36 = 3×12 → 成立 |
|
|
- 6年前の年齢関係：36-6=30 vs 12-6=6 → 30=5×6 → 成立 |
|
|
|
|
|
両方の条件が満たされているため、解は正しいと言えます。 |
|
|
</thinking> |
|
|
|
|
|
<answer> |
|
|
父親の年齢は36歳、息子の年齢は12歳です。 |
|
|
|
|
|
根拠： |
|
|
1. 現在の年齢関係：36 = 3×12（成立） |
|
|
2. 6年前の年齢関係：30 = 5×6（成立） |
|
|
3. 両方の条件を満たす唯一の解 |
|
|
</answer> |
|
|
``` |
|
|
|
|
|
-------------------------------------- |
|
|
## HOW TO USE |
|
|
------------------- |
|
|
### Setup |
|
|
``` |
|
|
pip install --upgrade transformers accelerate datasets trl |
|
|
``` |
|
|
------------------- |
|
|
### Predict(using AutoModelForCausalLM) |
|
|
```python |
|
|
|
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
|
|
|
|
|
|
|
|
model_name = "AXCXEPT/phi-4-open-R1-Distill-EZOv1" |
|
|
|
|
|
model = AutoModelForCausalLM.from_pretrained( |
|
|
model_name, |
|
|
torch_dtype="auto", |
|
|
device_map="auto" |
|
|
) |
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name) |
|
|
|
|
|
prompt = f""" |
|
|
ある父と子の年齢に関する問題があります。条件は次の通りです： |
|
|
|
|
|
父の現在の年齢は息子の現在の年齢の3倍です。 |
|
|
6年前、父の年齢は息子の年齢の5倍でした。 |
|
|
父と息子の年齢を求めてください。 |
|
|
""" |
|
|
messages = [ |
|
|
{"role": "system", "content": "You are an excellent AI."}, |
|
|
{"role": "user", "content": prompt} |
|
|
] |
|
|
text = tokenizer.apply_chat_template( |
|
|
messages, |
|
|
tokenize=False, |
|
|
add_generation_prompt=True |
|
|
) |
|
|
model_inputs = tokenizer([text], return_tensors="pt").to(model.device) |
|
|
|
|
|
generated_ids = model.generate( |
|
|
**model_inputs, |
|
|
max_new_tokens=2048 |
|
|
) |
|
|
generated_ids = [ |
|
|
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) |
|
|
] |
|
|
|
|
|
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] |
|
|
|
|
|
print(response) |
|
|
``` |
|
|
------------------- |
|
|
### Predict(using vllm) |
|
|
#### Setup |
|
|
``` |
|
|
pip install vllm |
|
|
vllm serve AXCXEPT/phi-4-open-R1-Distill-EZOv1 |
|
|
``` |
|
|
#### Predict |
|
|
``` |
|
|
from openai import OpenAI |
|
|
client = OpenAI( |
|
|
base_url="http://localhost:8000/v1", |
|
|
api_key="token-abc123", |
|
|
) |
|
|
|
|
|
prompt = f""" |
|
|
There is a question concerning the age of a father and his child. The conditions are as follows |
|
|
|
|
|
The father's current age is 3 times the son's current age. |
|
|
Six years ago, the father's age was five times the son's age. |
|
|
Find the ages of the father and the son. |
|
|
""" |
|
|
completion = client.chat.completions.create( |
|
|
model="AXCXEPT/EZO-phi-4-openr1-v1_917", |
|
|
messages=[ |
|
|
{"role": "system", "content": "You are an excellent AI. Please answer carefully and thoughtfully, in the same language as the instructions."} |
|
|
{"role": "user", "content": prompt} |
|
|
] |
|
|
) |
|
|
|
|
|
print(completion.choices[0].message.content) |
|
|
``` |
|
|
|
|
|
|
|
|
### Special Thanks |
|
|
Phi-4 develop team, open-r1 team developer, deepseek team, thanks for your special technology and knowledge. |
