How ColGrep-Agent-2B was trained: SFT on 27B teacher traces, then GRPO with F1-based rewards
IgorCarron · x · 2026-10-12
ColGrep-Agent-2B starts from MiniCPM5-2B and is trained in two stages: SFT on filtered code-search trajectories generated by Qwen3.5-27B, followed by GRPO reinforcement learning. The per-trajectory reward is ½ file F1 + ½ line F1, with a turn penalty after turn 2 and penalties for grep use, no search, and no finish — pushing the model to localize files accurately in fewer steps.
More from coding & agent
- Rigging in Blender via MCP: six-step recipe for Opus 5.5 to build and animate 3D scenes — sidahuj · 2026-10-12
- Diorama gives each agent its own git worktree — and bangs a giant gong when stuck — davidfromkansas · 2026-10-12
- Tracking people across home security cameras with cheap frame sampling and Haiku 5.5 — indragie · 2026-10-12
- How the cat dashboard works: Ring's AI tagging plus 1fps still frames fed to the model — indragie · 2026-10-12
- He built a cat activity dashboard with Opus 5.5 by sniffing Ring, Petlibro and Whisker app traffic — indragie · 2026-10-12
- Diorama breaks down agent work into abstraction layers so you can see what's happening — davidfromkansas · 2026-10-12