Z1-Coder
Unleashing the Reasoning Power of Large Language Models to Code Generation
News • Links • Getting Started • Introduction • Citation
News
- [2025/01/17] 🎉 We have released our Z1-Coder-1.5B and Z1-Coder-7B model and data through HuggingFace!
Links
- GitHub
- 🤗 Z1-Coder models
- 🤗 Z1-Coder data
Getting Started
We open source the code and scripts we used for data curation, training, and evaluation for Z1-Coder models, you can find more details in each directory.
src/eval
: Evaluation for Z1-Coder. To generate our training data, we use the QwQ-32B-Preview model. we curate reasoning trajectories on code-related datasets and propose self-invoking evolving to further refine models' reasoning behaviour in code generation.scr/train
: Training scripts for Z1-Coder. We train all the models with Fully Shard Data Parallel (FSDP) and set a global batch size to 1024 for 3 epochs using 2 NVIDIA A800-80G GPUs. We used greedy decoding for all results, with the maximum sequence length set to 1280. We use a learning rate of 5e-5 for the two training stages.
Introduction
To train Z1-Coder, we curate reasoning trajectories on code-related datasets and propose self-invoking evolving to further refine models' reasoning behaviour in code generation.
Model | Trajectory Dataset Download | Reference |
---|---|---|
SFT stage 1 Data | Z1Coder-Evol-CoT-110K | https://github.com/nlpxucan/WizardLM |
SFT stage 2 Data | Z1Coder-SelfInvoking-CoT-20K | https://github.com/CodeEval-Pro/CodeEval-Pro |
We fine-tune Qwen-2.5-Coder-Base (1.5B and 7B) for two stages with two trajectory datasets, yielding Z1-Coder-1.5B and Z1-Coder-7B respectively.
Z1-Coder-7B surpasses the best 7B code LLMs Qwen2.5-Coder-7B-Instruct, with only 1% its post-training data.
Model | Z1-Coder-7B | Qwen2.5-Coder-7B-Ins |
---|---|---|
Base Model | Qwen2.5-Coder-7B | Qwen2.5-Coder-7B |
SFT Data stage 1 | 110K (open-source) | 10M+ (open-source and in-house) |
SFT Data stage 2 | 20K (open-source) | 1M+ (in-house) |
Offline RL | No | DPO |
Citation
The code in this repository is mostly described in the post below. Please consider citing this work if you find the repository helpful.
@misc{z1-coder,
author = {Z1-Coder Team},
title = {Z1-Coder: Unleashing the Reasoning Power of Large Language Models to Code Generation},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/Z1-Coder/Z1-Coder}},
note = {Accessed: 2025-01-17},
year = {2025}
}
- Downloads last month
- 19