RL fine-tuning enables 4B models to surpass 235B in financial Q&A: Snorkel AI open-sources FinQA training environment

According to 1M AI News monitoring, Snorkel AI has released FinQA, a reinforcement learning training environment built on real SEC 10-K financial documents, and it is now open-sourced on the OpenEnv platform jointly maintained by Meta, PyTorch, and Hugging Face. FinQA covers 290 expert-labeled financial Q&A across 22 publicly traded companies (including Alphabet, Amazon, Apple, Bank of America, and Boeing), providing agents with 4 MCP tools: list available financial tables, retrieve table schemas, execute SQL queries, and submit answers. SQL strictly enforces filtering conditions and prohibits SELECT *, forcing the agent to retrieve only the required data instead of dumping an entire table.

Snorkel AI collaborated with the rLLM team at the University of California, Berkeley, using FinQA to perform reinforcement learning fine-tuning on Qwen3-4B. The results scored 59.7% on the financial Q&A benchmark SnorkelFinance, outperforming the same-series Qwen3-235B (51.37%). With parameter count at about 1/60 of the latter, inference costs dropped by about 90%. Key findings: large models can reason but produce hallucinated column names and ignore SQL constraints; after RL training, the smaller model can precisely call the tools instead—“tool discipline,” not scale, is the bottleneck.

FinQA is Snorkel AI’s first open-source environment released on OpenEnv, and future releases will roll out multi-turn enterprise environments covering industries such as healthcare, insurance, and legal.