GATS Cold-Start: Pure-OPD cold-start (ALFWorld)

Part of the Cold-start collection for the GATS paper (capability-aware on-policy distillation). Student: Qwen2.5-7B-Instruct; teacher: Qwen2.5-1.5B GRPO-trained on ALFWorld train tasks.

Cold-start phase: pure on-policy distillation (OPD)

The student rolls out on-policy in the environment and imitates the Qwen2.5-1.5B teacher (GRPO step-150 on ALFWorld train tasks) with a pure imitation objective (opd_only, coefficient 1.0 -- no RL reward term), for N steps where N is the GATS handoff step on this dataset: N = 52.

• Rollouts: 16 tasks/step, group n=8; lr 1e-6; N = 52 steps
• Checkpoint: global_step_52 (merged FSDP -> HF, bfloat16)
• Validation before any RL: 57.8% (train-time val at RL step 0; ALFWorld out-of-distribution split)

Intended use

Research on cold-start strategies for agent RL (imitation-first vs RL-from-scratch) and reproduction of the GATS paper baselines. Not a general-purpose assistant.

Provenance

Trained 2026-09-28 on the internal verl-based stack; checkpoints uploaded unmodified from the training runs (bfloat16 unless noted float32 above).