#539 · Primary category： AI Coding Assistants
AutoCoder
We introduced a new model designed for the Code generation task. Its test accuracy on the HumanEval base dataset surpasses that of GPT-4 Turbo (April 2024) and GPT-4o.
Project last updated：07/06/24
GitHub Stars
850
Forks
71
Contributors
2
License
Apache-2.0
Why we included this project
AutoCoder pairs a code-generation model with an interpreter that runs only when you ask it to verify the code, and it scores 90.9% on HumanEval, just ahead of GPT-4 Turbo and GPT-4o on the same set. Built on the DeepSeek-Coder and CodeQwen1.5 bases, it has a capability most open models lack: when you tell it to execute code, it can install the required packages on its own and keep iterating until it judges the result correct, sidestepping the dependency problems that normally stall local setup. The repository includes a Gradio demo of that interpreter flow and scripts to reproduce its published results across HumanEval, MBPP, and DS-1000, which helps you verify the claims before choosing a tool.
Articles for this project
No articles for this project yet.
To suggest a topic or contribute an article, contact us.
Related projects in this category
superpowers
An agentic skills framework & software development methodology that works.
ECC
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
skills
Skills for Real Engineers. Straight from my .agents directory.
opencode
The open source coding agent.
andrej-karpathy-skills
A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.