Stop Making Models Bigger, Make Them Behave — Kobie Crawdord, Snorkel

TL;DR

• 1

  A 4B model outperformed a 235B model on a finance tool-use task: Snorkel and UC Berkeley's RLLM team used RL to make a much smaller model behave better in a constrained FinQA environment.

• 2

  The failure mode was tool discipline, not intelligence: The 235B model guessed SQL against a non-existent table and then hallucinated, while the fine-tuned 4B model first called get_table_names, checked schema, and corrected its own query after an error.

• 3

  The training was cheap and fast: Crawford says the GRPO run took about 21 hours and cost less than $500 per run, making this a practical path for teams that want smaller, self-hosted models.

• 4

  Single-table training worked better than more complex curricula: Surprisingly, training only on single-table questions produced the strongest uplift, yet the gains generalized to harder multi-table FinQA Reasoning tasks too.

• 5

  Pass@1 roughly doubled after training: On the harder benchmark, performance jumped from 13.9 to 26.6, which Crawford frames as evidence that fixing the right behavior can matter more than scaling model size.

• 6

  Rubrics help find what behavior to train: Snorkel's broader point is to use detailed eval rubrics to locate the exact failure mode, then build targeted data and RL loops around that instead of defaulting to a bigger model.