A Reddit user is fine-tuning the Qwen 3.6 35B model using coherent fable data to improve local large language models. The project aims to gather extensive fable datasets from the community to enhance model performance.
- The author rented H200 GPUs to conduct the fine-tuning process.
- Early benchmarks show a +5% improvement in human evaluation and SWE scores.
- The user is requesting others to package their local fable data for inclusion in the dataset.
- The goal is to open-source the final model after further training runs.
The initiative highlights community-driven efforts to boost the capabilities of open-weight models through specialized data collection.