---
base_model:
- Qwen/Qwen3-Coder-30B-A3B-Instruct
library_name: transformers
pipeline_tag: text-generation
license: other
tags:
- code
- cpp
- parallel-computing
- parlaylib
---

# Model Card for qwen3-30b-sft-stage2-merged

This repository contains the model presented in the paper [ParEVO: Synthesizing Code for Irregular Data: High-Performance Parallelism through Agentic Evolution](https://huggingface.co/papers/2603.02510).

- **Project Website**: [https://quanquancliu.com/ParEVO/index.html](https://quanquancliu.com/ParEVO/index.html)
- **GitHub Repository**: [https://github.com/WildAlg/ParEVO](https://github.com/WildAlg/ParEVO)

## Model Details
- **Base Model:** `Qwen/Qwen3-Coder-30B-A3B-Instruct`
- **Model Type:** C++ Parallel Code Generation Model
- **Language:** C++
- **Parameters:** 30B

## Intended Use
This is a high-capacity large model fine-tuned for advanced parallel algorithmic reasoning on irregular data. Designed to target the **ParlayLib** library in C++, it produces efficient, natively parallel, and race-condition-free C++ code, surpassing general-purpose commercial models by focusing on concurrency and optimal scheduling.

## Training Data
Trained on the C++ subset of the **Parlay-Instruct Corpus** (comprising 13,120 training instructions) augmented by contrastive triplets to penalize synchronization breakdowns and performance anti-patterns. Training data can be found at the GitHub link: https://github.com/WildAlg/ParEVO

## Training Procedure
The model was aligned using a dual-stage pipeline on an NVIDIA H200 GPU:
- **Phase 1 (SFT):** Domain capability was established using QLoRA ($r=16, \alpha=32$) applied across all linear attention and MLP layers on standard DMOJ solutions and ParlayLib syntax.
- **Phase 2 (DPO):** Direct Preference Optimization (DPO) was applied to explicitly suppress failure modes (e.g., race conditions, deadlocks, and suboptimal span scaling). The model was trained on contrastive triplets coupling passing, performant code against failing or inefficient variants using a reduced learning rate of $5\text{e-}6$ and $\beta=0.1$.

## License

The ParEVO framework and datasets use a modular licensing structure to maximize open-source adoption, while the fine-tuned model weights inherit the license of their base model.

### 1. Model Weights License
The fine-tuned **`qwen3-parlay`** model weights are derivative works of `Qwen/Qwen3-Coder-30B-A3B-Instruct`. As such, the model weights and inference outputs are governed by the respective Qwen License (either [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0) or the [Tongyi Qianwen License Agreement](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT), depending on the exact base model release terms). Users must comply with the original use-case restrictions and terms set by Alibaba Cloud when using this model.

### 2. Software License (MIT License)
All software, scripts, the Evolutionary Coding Agent (ECA), and analysis code located in the [ParEVO repository](https://github.com/WildAlg/ParEVO) are licensed under the MIT License. Copyright (c) 2026 ParEVO Authors.

### 3. Dataset License (CC BY 4.0)
The Parlay-Instruct Corpus, ParEval evaluation trajectories, and DMOJ problem-solution datasets are licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

**Academic Citation:** If you use this model or the ParEVO evaluation framework in your research, please cite the repository at https://github.com/WildAlg/ParEVO and the following citation:

```bibtex
@inproceedings{yang2026parevo,
  title={ParEVO: Synthesizing Code for Irregular Data: High-Performance Parallelism through Agentic Evolution},
  author={Yang, Liu and Nie, Zeyu and Liu, Andrew and Zou, Felix and Altinb{\text{ü}}ken, Deniz and Yazdanbakhsh, Amir and Liu, Quanquan C.},
  booktitle={arXiv Preprint},
  year={2026}
}
```