Matryoshka NLA (Qwen2.5-7B L20)
Collection
An NLA trained with random-length truncation of the verbalizer's output, so the most important information comes first. Checkpoints, data, baseline. • 8 items • Updated
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.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support