This project uses Unsloth for fine-tuning. All training data is converted to OpenAI Harmony format before training, but there may be cases where the output format doesn't conform to the OpenAI Harmony specification.
Do you want to use pre-trained model?
You can download pre-trained data from HuggingFace.
This project implements a sophisticated multi-phase fine-tuning pipeline for the GPT-OSS-20B model, leveraging conversation data from multiple state-of-the-art AI models to create a balanced, high-performance language model optimized for:
Advanced Coding (via GPT-5.2-codex-max)
Complex Reasoning (via Claude 4.5 Opus and GPT-5.2 high reasoning)
Balanced General Intelligence (via Claude 4.5 Sonnet)
Why This Approach?
Traditional fine-tuning often suffers from:
Catastrophic forgetting when training on sequential datasets
Imbalanced capabilities from single-source training
