Bespoke Labs Post-Trains Open Model on a Code Repo: SFT+GRPO Yields 57pp Accuracy Lift

heghbalz · x · 2026-09-04

Bespoke Labs published a blueprint for post-training the open Inkling model to excel on a specific GitHub repository, using fontTools as the test case.

Pipeline: Curate repo-specific tasks via SWE-Smith techniques, SFT with trajectories from a strong teacher model, then RL (GRPO) on repository-specialized environments
Results: SFT alone gives a 52pp improvement on held-out fontTools evals; adding RL lifts total improvement to 57pp over the base model
Generalization: Matches base Inkling on Terminal-Bench 2.1 and SWE-Bench Lite while becoming up to 40% more token efficient
Training was done with Tinker API; authors say the method applies to any codebase

Original post →

Bespoke Labs Post-Trains Open Model on a Code Repo: SFT+GRPO Yields 57pp Accuracy Lift

More from coding & agent