GSPO-DeepSeek-R1-Distill-Qwen-1.5B

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

GSPO-DeepSeek-R1-Distill-Qwen-1.5B

(★ 18)

No description available.

GSPO-DeepSeek-R1-Distill-Qwen-1.5B Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Group Sequence Policy Optimization (GSPO) Implementation

[![Stars](https://img.shields.io/github/stars/vivekvar-dl/GSPO-DeepSeek-R1-Distill-Qwen-1.5B?style=for-the-badge)](https://github.com/vivekvar-dl/GSPO-DeepSeek-R1-Distill-Qwen-1.5B/stargazers)
[![Forks](https://img.shields.io/github/forks/vivekvar-dl/GSPO-DeepSeek-R1-Distill-Qwen-1.5B?style=for-the-badge)](https://github.com/vivekvar-dl/GSPO-DeepSeek-R1-Distill-Qwen-1.5B/network/members)
[![License](https://img.shields.io/badge/License-Apache%202.0-blue?style=for-the-badge)](https://raw.githubusercontent.com/vivekvar-dl/GSPO-DeepSeek-R1-Distill-Qwen-1.5B/main/LICENSE)
[![Python](https://img.shields.io/badge/Python-3.8+-blue?style=for-the-badge&logo=python)](https://www.python.org/)
[![PyTorch](https://img.shields.io/badge/PyTorch-2.0+-EE4C2C?style=for-the-badge&logo=pytorch)](https://pytorch.org/)
[![HuggingFace](https://img.shields.io/badge/%F0%9F%A4%97%20Model-vivekvar/GSPO--DeepSeek--R1--Distill--Qwen--1.5B-yellow?style=for-the-badge)](https://huggingface.co/vivekvar/GSPO-DeepSeek-R1-Distill-Qwen-1.5B)

**Open-source implementation of the GSPO algorithm by Zheng et al. (Qwen Team, Alibaba Inc.)** — optimized for NVIDIA H100 GPUs with comprehensive baseline comparisons (PPO, GRPO). Distills reasoning capabilities into smaller, deployable models.

## Paper Reference

**Group Sequence Policy Optimization**  
*Chujie Zheng, Shixuan Liu, Mingze Li, Xiong-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, Jingren Zhou, Junyang Lin*  
Qwen Team, Alibaba Inc.

## Quick Links

- **Trained Model**: [vivekvar/GSPO-DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/vivekvar/GSPO-DeepSeek-R1-Distill-Qwen-1.5B)
- **Training Logs**: [Wandb Run - GSPO Robust Training](https://wandb.ai/domainluther1234-usha-rama-college-of-engineering-and-te/gspo-robust-training/runs/pmyrt2ul/overview)
- **Validation Results**: See `validation_results/gspo_dynamics.png` for training dynamics visualization

## Algorithm Overview

Group Sequence Policy Optimization (GSPO) addresses limitations in existing policy optimization methods by introducing sequence-level importance ratios instead of token-level ratios. Key improvements include:

- **Sequence-level importance sampling**: Computes ratios over complete response sequences
- **Length normalization**: Applies normalization to handle variable sequence lengths
- **Enhanced stability**: Maintains training stability under high clipping rates (50-75%)
- **Superior performance**: Demonstrates improved performance on reasoning tasks

## Implementation Features

### Core Components
- **GSPOTrainer**: Main training class implementing the GSPO algorithm
- **GSPOConfig**: Configuration management for training parameters
- **Custom Dataset**: Specialized dataset for reasoning task evaluation
- **Baseline Comparisons**: PPO and GRPO implementations for scientific validation

### Technical Optimizations
- **H100 GPU optimization**: Memory-efficient training for large-scale models
- **8-bit optimizer support**: Reduced memory footprint using bitsandbytes
- **Gradient checkpointing**: Memory optimization for large models
- **Mixed precision training**: bfloat16 support for improved efficiency

## Performance Results

### Benchmark Performance
| Benchmark | Accuracy | Notes |
|-----------|----------|-------|
| ZebraLogic Reasoning | 60.0% | Logic puzzles and reasoning tasks |
| Custom Math Problems | 75.8% | Step-by-step mathematical reasoning |

### Algorithm Comparison
| Method | Reward Improvement | Clipping Stability | Training Stability |
|--------|-------------------|-------------------|-------------------|
| **GSPO** | **-1.4%** | **50-75%** | **Stable** |
| GRPO | -3.8% | 0.01% | Unstable |
| PPO | -2.9% | 0.02% | Degraded |

![Training Dynamics](https://raw.githubusercontent.com/vivekvar-dl/GSPO-DeepSeek-R1-Distill-Qwen-1.5B/main/validation_results/gspo_dynamics.png)

*Figure: GSPO training dynamics showing stable convergence and consistent performance improvements*

## Installation

### Prerequisites
- Python 3.8+
- CUDA-compatible GPU (recommended: H100, A100, or RTX 4090)
- PyTorch 2.0+

### Dependencies
```bash
# Install from requirements.txt
pip install -r requirements.txt

# Or minimal installation
pip install -r requirements_minimal.txt
```

### Package Installation
```bash
# Install as editable package
pip install -e .
```

## Usage

### Basic Training
```python
from gspo import GSPOTrainer, GSPOConfig, GSPOCustomDataset

# Initialize configuration
config = GSPOConfig(
    learning_rate=1e-7,
    left_clip_range=0.002,
    right_clip_range=0.002,
    group_size=4,
    batch_size=2
)

# Load model and tokenizer
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B")
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B")

# Initialize trainer
trainer = GSPOTrainer(model, tokenizer, config)

# Load dataset
dataset = GSPOCustomDataset()
train_data = dataset.generate_dataset(500)

# Training loop
for epoch in range(config.num_epochs):
    trainer.train_step(train_data)
```

### Pre-trained Model Usage
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load the trained model
model_name = "vivekvar/GSPO-DeepSeek-R1-Distill-Qwen-1.5B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)

# Generate response
prompt = "Solve step by step: If 3x + 7 = 22, what is x?"
inputs = tokenizer(prompt, return_tensors="pt")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.3,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(response)
```

## Repository Structure

```
gspo-implementation/
├── gspo/                    # Core GSPO implementation
│   ├── trainer.py          # Main GSPO trainer class
│   ├── config.py           # Configuration management
│   ├── dataset.py          # Custom dataset implementation
│   └── data_loader.py      # Data loading utilities
├── experiments/             # Experiment scripts
│   ├── train_gspo.py       # Main training script
│   ├── run_baseline_comparison.py  # PPO/GRPO comparison
│   ├── run_benchmarks.py   # Benchmark evaluation
│   └── monitor_realtime.py # Real-time training monitor
├── results/                 # Experimental results
│   ├── baselines/          # Baseline comparison results
│   ├── benchmarks/         # Benchmark evaluation results
│   └── evaluations/        # Model evaluation results
├── data/                   # Training datasets
├── docs/                   # Documentation
├── validation_results/     # Validation and training dynamics
└── wandb/                  # Training logs and metrics
```

## Experiments

### Training GSPO Model
```bash
python experiments/train_gspo.py \
    --model_name "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B" \
    --num_epochs 8 \
    --batch_size 2 \
    --learning_rate 1e-7 \
    --output_dir "./gspo_results"
```

### Baseline Comparison
```bash
python experiments/run_baseline_comparison.py \
    --algorithms "PPO,GRPO,GSPO" \
    --num_epochs 3 \
    --output_dir "./comparison_results"
```

### Benchmark Evaluation
```bash
python experiments/run_benchmarks.py \
    --model_path "./gspo_results/best_model" \
    --benchmarks "zebralogic,math,custom"
```

## Training Configuration

### Recommended Hyperparameters
- **Learning Rate**: 1e-7 (conservative for stability)
- **Clipping Range**: ±0.002 (tight bounds for sequence-level ratios)
- **Batch Size**: 2-4 (memory constraints on large models)
- **Group Size**: 4 (optimal for GSPO algorithm)
- **Optimizer**: AdamW with 8-bit optimization

### Hardware Requirements
- **Minimum**: 24GB VRAM (RTX 4090, A5000)
- **Recommended**: 80GB VRAM (H100, A100)
- **Training Time**: 4-8 hours for full training on H100

## Validation Results

The implementation has been thoroughly validated against the original paper specifications:

- **Algorithm Correctness**: Sequence-level importance ratios implemented as described
- **Performance Consistency**: Results align with paper benchmarks
- **Training Stability**: Demonstrates superior stability compared to token-level methods
- **Computational Efficiency**: H100-optimized implementation reduces training time

Detailed validation metrics and training dynamics are available in the `validation_results/` directory.

## Citation

If you use this implementation in your research, please cite both the original paper and this implementation:

### Original GSPO Paper
```bibtex
@article{zheng2024gspo,
  title={Group Sequence Policy Optimization},
  author={Chujie Zheng and Shixuan Liu and Mingze Li and Xiong-Hui Chen and Bowen Yu and Chang Gao and Kai Dang and Yuqiong Liu and Rui Men and An Yang and Jingren Zhou and Junyang Lin},
  journal={Qwen Team, Alibaba Inc.},
  year={2024}
}
```

### This Implementation
```bibtex
@misc{gspo_implementation2024,
  title={Open-Source Implementation of GSPO with H100 Optimization},
  author={Varikuti Sara Vivek},
  year={2025},
  note={Implementation of Zheng et al. GSPO algorithm},
  url={https://github.com/vivekvar-dl/GSPO-DeepSeek-R1-Distill-Qwen-1.5B}
}
```

## Contributing

Contributions are welcome! Please see our contributing guidelines for:
- Code style and formatting requirements
- Testing procedures and validation
- Documentation standards
- Pull request process

## License

This project is licensed under the Apache License 2.0 - see the [LICENSE](LICENSE) file for details.

## Acknowledgments

We thank the Qwen Team at Alibaba Inc. for developing the original GSPO algorithm and making their research publicly available. This implementation builds upon their foundational work to provide an accessible, optimized version for the research community.

**Original Authors**: Chujie Zheng, Shixuan Liu, Mingze Li, Xiong-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, Jingren Zhou, Junyang Lin

## Support

For questions about the implementation:
- **Issues:** [GitHub Issues](https://github.com/vivekvar-dl/GSPO-DeepSeek-R1-Distill-Qwen-1.5B/issues)
- **Docs:** Check the `docs/` directory
- **Training Logs:** [Wandb Run](https://wandb.ai/domainluther1234-usha-rama-college-of-engineering-and-te/gspo-robust-training/runs/pmyrt2ul/overview)

For questions about the original algorithm, please refer to the paper by Zheng et al. and the Qwen Team at Alibaba Inc.

---

**⭐ Star this repo if you find it useful!**

**Built by [Vivek Varikuti](https://vivekvari.dev)**

GSPO-DeepSeek-R1-Distill-Qwen-1.5B

GSPO-DeepSeek-R1-Distill-Qwen-1.5B Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

GSPO-DeepSeek-R1-Distill-Qwen-1.5B

GSPO-DeepSeek-R1-Distill-Qwen-1.5B Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README