EmberJin/Minecraft-Agent-CoT-Qwen3.5-9B-v2

Minecraft Agent CoT v2 (Qwen3.5-9B, checkpoint 520 / epoch 4)

CoT continue-SFT model with cleaned thought annotations (blacklist-filtered + perception-only V2 prompt + Combat rebalanced to 23%). Current best CoT model.

Training

• Base: EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3 (24.9% easy-ng, h29)
• Data: 8,261 rows = v1 rows with 1,452 toxic thoughts stripped (hallucinated outcomes, wait-on-noop) + 1,600 newly Combat-annotated trajectories (V2 prompt, 99.6% pass rate) + replay
• LR 1e-6, 5 epochs, epoch-4 checkpoint (520 steps; epoch 5 lost to a crash at save time)

Results (easy-ng, h29, 3 rollouts, seed 42)

	Overall	Embodied	Combat
Base (no CoT)	24.9%	24.8%	25.2%
v1 (raw thoughts, e4)	25.3%	27.5%	18.4%
This (v2-e4, cleaned)	29.2%	31.2%	23.1%
No-thought control (e5)	31.5%	32.0%	29.9%

For rollout QC — what to compare

1. vs v1 (link below): hallucinated completion claims should be largely gone; Combat partially recovered (+4.7pp). Remaining gap to the no-thought control (~2pp at matched steps) is the open question — is it residual thought interference with the action distribution?
2. Thought emission rate: still far below the 12% training trigger rate — check whether thoughts now appear at the right moments (approach/initiation) vs v1.
3. Combat residuals: attack-spam with frozen camera was v1's signature failure; check if reduced but not eliminated.

A no-thought control (same data, thoughts stripped) reached 31.5% — i.e. current cleaned thoughts still do not beat pure data exposure. The interesting residual question for QC: do v2 thoughts help conditional on appearing, even if too rare to move the aggregate?

Serving

vllm serve . --served-name cot-v2-e4 \
  --max-model-len 32768 --limit-mm-per-prompt image=30

Output format: Thought: <perception> | <decision>\nAction: move(dx, dy) and press(keys) [and click(left|right)].