100 lines
2.9 KiB
Markdown
100 lines
2.9 KiB
Markdown
|
---
|
|||
|
|
base_model: unsloth/Qwen2.5-0.5B-Instruct
|
|||
|
|
library_name: transformers
|
|||
|
|
license: apache-2.0
|
|||
|
|
datasets:
|
|||
|
|
- HoangHa/pensez-grpo
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
tags:
|
|||
|
|
- math
|
|||
|
|
- trl
|
|||
|
|
- unsloth
|
|||
|
|
- grpo
|
|||
|
|
- transformers
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Model Card for Math-RL
|
|||
|
|
|
|||
|
|
## Model Details
|
|||
|
|
|
|||
|
|
This model is a fine-tuned version of Qwen2.5-0.5B-Instruct, optimized with Group Relative Policy Optimization (GRPO) on a curated math dataset of 700 problems.
|
|||
|
|
The fine-tuning process aims to enhance the model’s step-by-step reasoning ability in mathematical problem solving, improving its performance on structured reasoning tasks.
|
|||
|
|
|
|||
|
|
### Model Description
|
|||
|
|
|
|||
|
|
|
|||
|
|
- **Language(s) (NLP):** English
|
|||
|
|
- **License:** MIT
|
|||
|
|
- **Finetuned from model:** Qwen2.5-0.5B-Instruct
|
|||
|
|
- **Fine-tuning Method**: GRPO with LoRa
|
|||
|
|
- **Domain**: Mathematics (problem-solving, reasoning)
|
|||
|
|
- **Dataset Size**: ~700 examples
|
|||
|
|
|
|||
|
|
|
|||
|
|
## Uses
|
|||
|
|
|
|||
|
|
### Direct Use
|
|||
|
|
|
|||
|
|
The model is intended for:
|
|||
|
|
|
|||
|
|
- Educational purposes: assisting students with math problems
|
|||
|
|
- Research on small-scale RLHF-style fine-tuning (GRPO)
|
|||
|
|
- Experiments in reasoning with small instruction-tuned models
|
|||
|
|
- Serving as a lightweight math reasoning assistant in constrained environments
|
|||
|
|
|
|||
|
|
|
|||
|
|
## Bias, Risks, and Limitations
|
|||
|
|
|
|||
|
|
- Small Dataset: Fine-tuned only on 700 math problems, so generalization is limited.
|
|||
|
|
- Reasoning Errors: May produce incorrect or hallucinated answers. Always verify results.
|
|||
|
|
- Not a Math Oracle: Should not be used in high-stakes scenarios (e.g., exams, grading, critical calculations).
|
|||
|
|
- Limited Scope: Performance is strongest on problems similar to the fine-tuning dataset; outside domains may degrade.
|
|||
|
|
- Language: While the base model supports multiple languages, math-specific fine-tuning was primarily English-based.
|
|||
|
|
|
|||
|
|
|
|||
|
|
## How to Get Started with the Model
|
|||
|
|
|
|||
|
|
Use the code below to get started with the model.
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
|
|||
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|||
|
|
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained("khazarai/Math-RL")
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|||
|
|
"khazarai/Math-RL",
|
|||
|
|
device_map={"": 0}
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
question = """
|
|||
|
|
Translate the graph of the function $y=\sin 2x$ along the $x$-axis to the left by $\dfrac{\pi }{6}$ units, and stretch the ordinate to twice its original length (the abscissa remains unchanged) to obtain the graph of the function $y=f(x)$. If the minimum value of the function $y=f(x)+a$ on the interval $\left[ 0,\dfrac{\pi }{2} \right]$ is $\sqrt{3}$, then $a=\boxed{\_\_\_\_\_}$.
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
system = """
|
|||
|
|
Respond in the following format:
|
|||
|
|
<reasoning>
|
|||
|
|
...
|
|||
|
|
</reasoning>
|
|||
|
|
<answer>
|
|||
|
|
...
|
|||
|
|
</answer>
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
messages = [
|
|||
|
|
{"role" : "system", "content" : system},
|
|||
|
|
{"role" : "user", "content" : question}
|
|||
|
|
]
|
|||
|
|
text = tokenizer.apply_chat_template(
|
|||
|
|
messages,
|
|||
|
|
tokenize = False,
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
from transformers import TextStreamer
|
|||
|
|
_ = model.generate(
|
|||
|
|
**tokenizer(text, return_tensors = "pt").to("cuda"),
|
|||
|
|
max_new_tokens = 2048,
|
|||
|
|
streamer = TextStreamer(tokenizer, skip_prompt = True),
|
|||
|
|
)
|
|||
|
|
```
|