Update README.md
Browse files
README.md
CHANGED
@@ -36,6 +36,40 @@ evaluation rubrics, and a score along with the corresponding reasoning.
|
|
36 |
- **Repository:** https://github.com/rubricreward/r3
|
37 |
- **Paper:** https://arxiv.org/abs/2505.13388
|
38 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
## License and use
|
40 |
|
41 |
R3 is licensed under the Apache 2.0 license.
|
|
|
36 |
- **Repository:** https://github.com/rubricreward/r3
|
37 |
- **Paper:** https://arxiv.org/abs/2505.13388
|
38 |
|
39 |
+
## Using the Model
|
40 |
+
|
41 |
+
|
42 |
+
```python
|
43 |
+
from transformers import AutoTokenizer
|
44 |
+
from vllm import LLM, SamplingParams
|
45 |
+
|
46 |
+
model_path = "rubricreward/R3-Qwen3-4B-LoRA-4k"
|
47 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
48 |
+
sampling_params = SamplingParams(temperature=0.6, top_p=0.95, max_tokens=8192, min_p=0, top_k=20)
|
49 |
+
|
50 |
+
llm = LLM(
|
51 |
+
model=model_path,
|
52 |
+
dtype="bfloat16",
|
53 |
+
max_model_len=10000,
|
54 |
+
tensor_parallel_size=2,
|
55 |
+
gpu_memory_utilization=0.9,
|
56 |
+
enforce_eager=True,
|
57 |
+
)
|
58 |
+
|
59 |
+
messages: list[dict[str, str]] = [
|
60 |
+
{'content': "Evaluate the response based on the given task, input, response, and evaluation rubric. Provide a fair and detailed assessment following the rubric...", 'role': 'user'}
|
61 |
+
]
|
62 |
+
|
63 |
+
list_text = tokenizer.apply_chat_template(
|
64 |
+
messages,
|
65 |
+
tokenize=False,
|
66 |
+
add_generation_prompt=True,
|
67 |
+
enable_thinking=True # Switch between thinking and non-thinking modes.
|
68 |
+
)
|
69 |
+
|
70 |
+
outputs = llm.generate(list_text, sampling_params)
|
71 |
+
```
|
72 |
+
|
73 |
## License and use
|
74 |
|
75 |
R3 is licensed under the Apache 2.0 license.
|