Stop Making Models Bigger, Make Them Behave
summary
A 235B Qwen 3 model queries a non‑existent table twice then hallucinates a YouTube ads growth number; a 4B model finetuned with RL uses get_table_name, then get_table_info, runs a corrected SQL query after a column error, and returns the right answer. Training runs cost under $500 and use a FinQA environment; single‑table RL training lifts multi‑table FinQA pass@1 from 13.9%→26.6%, showing concrete tool‑use fixes beat bigger models for this task.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 3:45Snorkel’s mission
- 6:25Bigger models aren’t always better
- 9:11High-quality datasets with experts
- 11:16Reinforcement learning setup
- 13:42Evaluation results
- 16:30Key behaviors for success
- 18:35Rubrics for model weaknesses
- 20:01Summary
related talks
15mOlmo-Thinking: Training a Fully Open Reasoning Model
29mHow AI Models Get Smarter
51mDiscover BigQuery ML, build your own CREATE MODEL stmt
33mThe Right Model for the Right Job: Cutting LLM Costs Without Cutting Quality
35mImproving Credibility of ML Solutions
28mFrom Tables to Answers: building QA System for In-Document Searches
42mPost-Training Small Language Models: the adventures of a practitioner
21mFrom Data to Precision: Fine-Tuning LLMs for Proprietary Query Languages