← back
Stop Making Models Bigger, Make Them Behave — Kobie Crawford, Snorkel
Takeaway
Train reliable domain-specific tool behavior before paying for broader model capacity that may not solve the actual failure.
Summary
- Snorkel and Berkeley's RLLM team study using reinforcement learning to make a 4-billion-parameter model outperform a 235-billion-parameter model on financial tool-use tasks.
- The approach targets behavior with high-quality, expert-informed data rather than assuming a larger model is always the answer.
- Smaller models can support lower costs, faster inference, and deployment requirements around on-premises data control.
- A large-model failure example queries nonexistent tables twice and then invents an answer instead of inspecting available data sources.
reinforcement-learningsmall-modelstool-use
Original description
Qwen 3 235B was asked for YouTube's year over year ad revenue growth from 2023 to 2024. It queried a table that didn't exist, tried again, got nothing back both times, and hallucinated an answer. The 4B model Snorkel finetuned with RL called `get_table_name` first, inspected the schema, ran a query, hit a column error, self-corrected, and got the right answer. The training run cost under $500. Kobe Crawford covers why tool discipline matters more than reasoning depth for this class of tasks, how single table training transferred cleanly to harder multi table problems (13.9% to 26.6% on the FinQA reasoning benchmark), and why breaking evals into rubrics helps identify which specific behavior to fix before writing any training data. Speaker info: / kobie-crawford https://snorkel.ai/author/kobie-crawf...