RL Fine-Tuning Takes 4B Model Finance QA Beyond 235B: 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, now open-sourced on the OpenEnv platform jointly maintained by Meta PyTorch and Hugging Face. FinQA covers 290 expert-annotated financial Q&A from 22 public companies (including Alphabet, Amazon, Apple, Bank of America, Boeing), providing the Agent with 4 MCP tools: list available financial tables, get table schema, execute SQL queries, submit answers. SQL mandates filter conditions and prohibits `SELECT *`, forcing the Agent to only fetch necessary data rather than dump the entire table.

Snorkel AI collaborated with the rLLM team at the University of California, Berkeley to fine-tune Qwen3-4B through FinQA for reinforcement learning, achieving a score of 59.7% on the SnorkelFinance financial Q&A benchmark, surpassing its sibling Qwen3-235B (51.37%), with around 1/60th of the parameters and approximately 90% reduced inference cost. Key findings: large models can infer but may suffer from illusionary column names and ignore SQL constraints; post-RL training, smaller models can more accurately invoke tools, where "tool discipline" rather than scale is the bottleneck.

FinQA marks Snorkel AI's first open-source environment in OpenEnv, with upcoming releases of multi-turn enterprise environments covering industries such as healthcare, insurance, and law.