Typicality reward function#17

Merged

fingriffin merged 14 commits into

mainfrom

feature/typicality-reward-func

Jun 22, 2026

Commits on Jun 16, 2026

feat(rl): add grpo transform factory and reward func stub, wire in rl hyperparams
Fin Griffin
committed
feat(docs): add first draft for reward function
Fin Griffin
committed
feat(rl): add style embedding based reward function
Fin Griffin
committed
fix(rl): fix wiring of nested trl parameters
Fin Griffin
committed
fix: remove single column guard on dataset preprocessing
Fin Griffin
committed
fix: relative import error for grpo
Fin Griffin
committed

Commits on Jun 17, 2026

fix: force uv-managed python via .toml
Fin Griffin
committed
fix: monkeypatch GRPOTrainer chat template to use enable_thinking=False when passed into config
Fin Griffin
committed
fix(docs): clarify that reference model, wrt to reward function, is base model not true GRPO reference model
Fin Griffin
committed

Commits on Jun 18, 2026

feat(docs): add first draft of second reward function
Fin Griffin
committed
feat(rewards): add typicality reward function
Fin Griffin
committed

Commits on Jun 22, 2026

chore(grpo): remove style embedding based reward
Fin Griffin
committed
feat(docs): remove style embedding reward from documentation
Fin Griffin
committed
test(grpo): add test coverage for GRPO utilities
Fin Griffin
committed