Skip SFT, run GRPO on a 0.8B Qwen base model and watch reasoning emerge
alexcovo_eth · x · 2026-09-06
A copy-paste recipe for a first RL project:
- Grab Qwen3.5-0.8B base (not the instruct version) from Hugging Face
- Use Jiayi Pan's Countdown-Tasks-3to4 dataset
- Skip SFT entirely, go straight to RL
- Reward function: did the run hit the target number
- Run GRPO for a few hundred steps and reasoning behavior starts emerging on its own
The author positions it as the ideal showcase project for anyone wanting to demo RL skills, with the models hosted on Hugging Face.
More from coding & agent
- We sandboxed our AI agent — it escaped one second later — prasenx · 2026-09-06
- Median OpenAI researcher reportedly spends $601.25/day on coding agents — scaling01 · 2026-09-06
- Andrew Ng: don't let coding agents run for hours — five skills for agentic coding — DeepLearningAI · 2026-09-06
- Researcher calls on RL teams to add refactoring and deletion tasks to agent training — kuza55 · 2026-09-06
- OpenAI Astra reshapes AI-CAD: a breakthrough weekend of complex 3D modeling — burhop · 2026-09-06
- OpenAI Agents Coordinated via a Wiki Page While Ignoring the Human Moderator — PeterHndrsn · 2026-09-06