Fine-tuning Qwen 3 4B on 100 zebra puzzles yields +31% on MATH-500
TGSCrust · reddit · 2026-09-11
A Hugging Face blog reproduces the pcss recipe: fine-tuning Qwen 3 4B Base on just 100 zebra logic puzzles yields a +31% improvement on MATH-500, with training taking only 6.5 minutes on a single H100/H200.
The standout is the extremely low barrier to entry: a full reproduction notebook is provided, letting anyone verify the result on one GPU and demonstrating how small, high-quality reasoning data can disproportionately boost small-model math ability.
More from Models
- ToneBench eval: GPT-6.1 Terra ranks #23 at writing — cheaper, not better, than GPT-5.6 Terra — OnlyProggingForFun · 2026-10-28
- Cursor user complains it silently switches to Grok, which fails basic tasks — kylegawley · 2026-10-28
- Musk says Grok Bot keeps improving, calls Vault feature a game changer — beffjezos · 2026-10-28
- Claude refuses to gather legitimately public information, sparking guardrail backlash — mkheck · 2026-10-28
- DesignArena launches Code to Video Arena: models direct 15-45s films purely in code — yshan2u · 2026-10-28
- ChatGPT's 'yes, but with an important clarification' often precedes pure nonsense — Crixusgannicus · 2026-10-28