EmberJin/Minecraft-Agent-CoT-Qwen3.5-9B-v1

Minecraft Agent CoT v1 (Qwen3.5-9B, checkpoint 420 / epoch 4)

CoT continue-SFT model with raw (uncleaned) thought annotations. This is the problem specimen version: kept for rollout quality inspection and failure analysis of the original CoT annotation approach.

Training

• Base: EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3 (24.9% easy-ng, h29)
• Data: 6,661 rows = 5,329 CoT-annotated trajectories (gemini-3.8-flash, old prompt) + 1,332 replay
• LR 1e-6, 5 epochs, epoch-4 checkpoint (420 steps)

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

	Overall	Embodied	Combat
Base (no CoT)	24.9%	24.8%	25.2%
This (v1-e4)	25.3%	27.5%	18.4%

⚠️ Known issues to look for during rollout QC

1. Hallucinated completion claims (most important): thoughts asserting unverifiable outcomes — "task complete", "has been killed", "taking damage", "within attack range" — that contradict the actual game state. Reference failure: an episode where the model sheared a sheep, then narrated "the task to kill the sheep is complete" twice, and looped until timeout.
2. Attack-spam with frozen camera: in Combat, wrong "in range" assertions replayed in the history window lock the model into swinging at air (~170 attacks, camera delta ≈ 0.1/step) without re-aiming.
3. Low thought emission: only 0.3–1.5% of steps emit a thought (training data had 12% at decision points). The format is learned perfectly; the when-to-think trigger is not.
4. Combat degradation (-6.8pp vs base) — attributed to the above thought pathologies, not data mix (verified by a no-thought control).

A companion model ...-v2 was trained on cleaned annotations (blacklist + rewritten prompt) — compare rollouts between the two to isolate annotation-quality effects.

Serving

vllm serve . --served-name cot-v1-e4 \
  --max-model-len 32768 --limit-mm-per-prompt image=30

Output format: Thought: <perception> | <state> | <decision>\nAction: move(dx, dy) and press(keys) [and click(left|right)]. Parse actions from the Action: prefix; thoughts appear at ~1% of steps.