OpenCoder SFT Stage 2
Corpus dedicated to training code generation models, combining educational data, Python documentation and validated synthetic instructions.
Approximately 300,000 examples divided into 4 subsets in JSON format (instruction, code, test case)
MIT
Description
The dataset OpenCoder SFT Stage 2 is an open-source resource for the supervised refinement of generative language models specialized in programming. It includes four subsets: educational instructions with validated tests, prompts from MagicCoder and McEval, and a corpus generated from Python documentation. Each entry contains a clear instruction, associated Python code, and sometimes a test case, allowing for robust and controlled learning.
What is this dataset for?
- Train models capable of generating accurate and functional Python code
- Evaluate the algorithmic understanding and the ability of an LLM to respond to technical prompts
- Improving SFT Performance for Development-Oriented Conversational Agents
Can it be enriched or improved?
Yes, this dataset is completely reusable and extensible. You can add other programming languages, translate instructions, introduce additional metadata (difficulty level, domain, code structure) or apply human annotation to refine the quality of the pairs (instruction/code/test).
🔎 In summary
🧠 Recommended for
- Coding LLM Developers
- Technical NLP researchers
- AI/Code Bootcamps
🔧 Compatible tools
- Hugging Face Transformers
- PyTorch
- DeepSpeed
- LoRa
- VllM
💡 Tip
For effective training, start with a subset like educational_instruct Before Expanding to the Full Set
Frequently Asked Questions
Does this dataset only contain Python code?
Yes, all data is centered on Python, in particular via pydoc, compiler-validated tests and adapted instructions.
Is it suitable for training a developer chat agent?
Perfectly. It covers realistic cases of use, explanation of functions, and the generation of testable code.
Can this dataset be combined with RLHF or advanced refinement methods?
Yes, several subsets like educational_instruct and mceval_instruct are already optimized for RLHF approaches.
