[Model Release] I trained a 9B model to be agentic Data Analyst (Qwen3.5-9B + LoRA). Base model failed 100%, this LoRA completes 89% of workflows without human intervention.

An independent developer has demonstrated that small language models can achieve true agentic autonomy through specialized training, not just complex prompting frameworks. By creating a LoRA (Low-Rank Adaptation) adapter called CoPaw-Flash-9B-DataAnalyst-LoRA for the Qwen3.5-9B model, the developer transformed a model that previously failed 100% of data analysis workflows into one that completes 89.7% autonomously. The key innovation was training on massive, multi-step trace datasets covering real-world scenarios rather than standard instruction tuning.

The adapter enables the 9-billion parameter model to plan, execute Python code, debug, visualize data, and summarize findings in continuous loops up to 50 iterations without human prompting. In benchmark tests on 29 real Kaggle datasets, the base model averaged just 1.2 iterations before stopping, while the LoRA-enhanced version averaged 26 autonomous iterations, writing code and generating charts until tasks were complete. The system runs locally on consumer hardware, requiring only 6-24GB VRAM depending on quantization.

The developer has released weights on Hugging Face and an inference framework called data-analyst to handle the tool-calling loop, along with a live demo. This proof-of-concept demonstrates that small models can serve as capable autonomous agents when trained on complete workflow scenarios rather than isolated tasks. The developer is now seeking compute sponsorship to expand the methodology to coding assistants and research agents.