Final runs, live. LFM2.5-24B-A2B trains for 200 steps, LFM3-1.7B for 400 and the TwIL-8B experiment for 100, each on SmolDataEnvs and 230 traj2envs tasks. Charts update as the steps and evals land.

Runs

one LoRA adapter per model and harness; each model's three trainers share one vLLM. Held-out columns show the best checkpoint (@step): highest pass@1, and fewest tool calls per solved task, against the base

Train reward

Held-out pass@1

scheduled eval of every saved checkpoint under its own harness; dashed: the base model

Tool calls per correct answer

lower is better: how many tool calls it takes to get a task right. Failed attempts are left out, and the reward's tool bonus pushes this down. Dashed: the base model

Timeline