---
license: apache-2.0
tags: [nla, qwen2.5, checkpoints]
---
# v3 RL resumable checkpoint (iter_0000200)
Full actor DCP (+optimizer state) and critic HF export from the v3 truncation-RL run
(KL 0.03, uniform token truncation ~U[1,120], bullets prompt). To resume: place
actor_dcp/ at $RUN_DIR/actor/iter_0000200, critic_hf/ at $RUN_DIR/critic/iter_0000200/hf,
write '200' to $RUN_DIR/actor/latest_checkpointed_iteration.txt, run run_rl_v3.sh.
Inference-ready per-iter models live in syvb/nla-qwen2.5-7b-L20-v3-rl.