Qwen3-4B-GRPO-KL-math-reasoning/README.md

---
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen3-4B
tags:
- reinforcement-learning
- grpo
- math-reasoning
- pipelinerl
datasets:
- gsm8k_train
- math_train
pipeline_tag: text-generation
---

# Qwen3-4B-GRPO-KL-math-reasoning

This model is a fine-tuned version of [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) using **GRPO (Group Relative Policy Optimization) with KL penalty** for mathematical reasoning.

Trained with [PipelineRL](https://github.com/ServiceNow/PipelineRL).

## Training Details

### Datasets

| Split | Datasets |
|-------|----------|
| Train | `gsm8k_train`, `math_train` |
| Test  | `gsm8k_test`, `math_500` |

### RL Algorithm

| Parameter | Value |
|-----------|-------|
| Algorithm | GRPO (Group Relative Policy Optimization) |
| Policy Loss | `ppo` |
| KL Coefficient | `0.001` |
| Epsilon (clip) | `0.02` |
| Divide Advantage by Std | `False` |
| Filter Zero Advantage Groups | `False` |
| Rollouts per Problem | `16` |

### Training Hyperparameters

| Parameter | Value |
|-----------|-------|
| Base Model | `Qwen/Qwen3-4B` |
| Learning Rate | `1e-06` |
| LR Scheduler | `cosine` |
| Warmup Steps | `25` |
| Max Training Steps | `1500` |
| Micro Batch Size | `2` |
| Gradient Accumulation | `128` |
| Effective Batch Size | `256` |
| Sequence Length | `8192` |
| Gradient Clipping | `0.3` |
| Weight Decay | `0.01` |
| Optimizer | `adamw_torch` |
| Precision | `bf16` |
| DeepSpeed | ZeRO Stage 3 |

## Evaluation Results

Pass@k on math reasoning benchmarks (N=32 samples per problem, temperature=1.0):

| Dataset | pass@1 | pass@2 | pass@4 | pass@8 | pass@16 | pass@32 |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| GSM8K (test) | 89.47 | 92.04 | 93.66 | 94.79 | 95.57 | 96.13 |
| MATH-500 | 81.04 | 86.44 | 90.03 | 92.55 | 94.46 | 96.00 |
| **Overall** | **87.15** | **90.50** | **92.66** | **94.17** | **95.26** | **96.10** |

*GSM8K test: 1319 problems · MATH-500: 500 problems · Overall: 1819 problems (overall weighted by problem count).*

## Training Curves

![Training Metrics](training_metrics.png)

## W&B Run

Full training logs: [https://wandb.ai/jaygala24-team/rl-post-training/runs/qwen3_4b_grpo_with_kl_2a1p1f_4xh100_197342_finetune_d0a43ea2](https://wandb.ai/jaygala24-team/rl-post-training/runs/qwen3_4b_grpo_with_kl_2a1p1f_4xh100_197342_finetune_d0a43ea2)

## Usage

### Transformers
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("jaygala24/Qwen3-4B-GRPO-KL-math-reasoning", revision="step-0200")  # or whatever branch name, e.g. "step-0400", "step-0600"
tokenizer = AutoTokenizer.from_pretrained("jaygala24/Qwen3-4B-GRPO-KL-math-reasoning", revision="step-0200")  # or whatever branch name, e.g. "step-0400", "step-0600"

prompt = "Please reason step by step, and put your final answer within \\boxed{{}}.\n\nWhat is the sum of 123 and 456?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=4096, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### vLLM

```python
from vllm import LLM, SamplingParams

llm = LLM(model="jaygala24/Qwen3-4B-GRPO-KL-math-reasoning", revision="step-0200")  # or whatever branch name, e.g. "step-0400", "step-0600"
sampling_params = SamplingParams(temperature=0.7, max_tokens=4096)

prompt = "Please reason step by step, and put your final answer within \\boxed{}.\n\nWhat is the sum of 123 and 456?"
outputs = llm.generate([prompt], sampling_params)
print(outputs[0].outputs[0].text)
```

## Framework

- [PipelineRL](https://github.com/ServiceNow/PipelineRL)
- [Transformers](https://github.com/huggingface/transformers)
- [DeepSpeed](https://github.com/microsoft/DeepSpeed) (ZeRO Stage 3)