RL Fine-Tuning Enables 4B Model to Outperform 235B in Financial Q&A: Snorkel AI Releases Open Source FinQA Training Environment

According to monitoring by 1M AI News, Snorkel AI has released FinQA, a reinforcement learning training environment built on real SEC 10-K financial documents, now open-sourced on the OpenEnv platform jointly maintained by Meta PyTorch and Hugging Face. FinQA covers 290 expert-annotated financial questions from 22 publicly traded companies, including Alphabet, Amazon, Apple, Bank of America, and Boeing, providing the Agent with four MCP tools: listing available financial tables, retrieving table structures, executing SQL queries, and submitting answers. SQL enforces filtering conditions and prohibits SELECT *, forcing the Agent to only retrieve the necessary data instead of dumping the entire table. Snorkel AI collaborated with the rLLM team at the University of California, Berkeley, to fine-tune Qwen3-4B using FinQA, resulting in a score of 59.7% on the financial Q&A benchmark SnorkelFinance, surpassing the same series Qwen3-235B (51.37%), with approximately 1/60th the number of parameters and a 90% reduction in inference cost. Key findings: while large models can reason, they may produce hallucinated column names and ignore SQL constraints; in contrast, the smaller model trained with RL can accurately invoke tools, indicating that ‘tool discipline’ rather than scale is the bottleneck. FinQA is the first open-source environment released by Snorkel AI on OpenEnv, with plans to launch multi-turn enterprise environments covering industries such as healthcare, insurance, and law in the future.