SlowGuess/qwen38-minecraft-r4

Qwen3.8-27B Minecraft per-tick action model (run "R4")

Qwen/Qwen3.8-27B, fine-tuned to play Minecraft from first-person screenshots. Each step outputs one native keyboard/mouse action (one 50 ms tick) through a play tool call.

The weights are the trainer's final save at step 4,041 (FP32), uploaded unchanged.

Training

Item	Setting
Base	Qwen/Qwen3.8-27B, with no Minecraft vision-language pre-training stage
Data	m2_onepf_book/seq/train.jsonl from SlowGuess/qwen38-minecraft-m2-data: 258,621 sequences, 6.17 M supervised actions
Sequence	1–30 consecutive ticks of one recording; task text and first frame, then one play call and its resulting frame per tick; every tick supervised; at most 16,384 tokens
Parameters	All trained. Vision encoder at 0.7× the language-model learning rate
Optimizer	AdamW, lr 5e-6, cosine schedule, 200 warmup steps, global batch 64, one epoch (4,041 steps)
Hardware	4×B200, FSDP, FP32 master weights with BF16 compute
Validation loss	0.0879 at step 4,000, on m2_repaired/seq/valid.jsonl

Usage

• System prompt and tool: the system prompt and the play tool definition used in training are in qwen38_policy_prompt.json.
• Turn format:
  • The user turn holds the task text and the current screenshot.
  • The model answers with one play call, {"action": {...}}.
  • The tool output is the next screenshot.
• Action format:
  • camera is [pitch_delta, yaw_delta] in degrees.
  • Other keys are 0/1, and omitted keys are 0.
• Context restart: training sequences hold at most 30 actions, so restart the context from the current frame after 30 actions.
• Settings we evaluated with:
  • thinking disabled (enable_thinking=False);
  • temperature 1.0, top_p 0.95, top_k 20;
  • vLLM with --dtype bfloat16.

Evaluation

Full commands, code versions (md5), training logs and every rollout (logs, per-step actions, model requests and responses, videos, raw frames) are in run/.

MineExplorer, fixed 20-scene subset (1,800 steps)

Hops	1	2	3	4	All
Success	4/6	2/6	1/4	0/4	7/20

Milestones 20/46 (macro MSR 49.2).

OpenHA tasks, 30 tasks × 3 runs, 600 steps

This is a diagnostic subset (10 mining, 10 crafting, 10 combat tasks), not the official OpenHA benchmark.

Category	Successes
Mine	14/30
Craft	23/30
Kill	9/30
All	46/90

In combat tasks, 46 % of the model's actions are empty, often while the target is in view.

200 ms action-chunk variant (chunk200/)

chunk200/ holds a second model, trained the same way as R4 from the same frames and actions. It outputs one 200 ms action chunk (4 ticks) per call instead of one tick, following Lumine and Scaling in Games. Its weights, prompt file, training logs and all rollouts are in that folder. On the same 175 MineExplorer scenes it succeeds on 75 scenes against R4's 61. On the OpenHA 30×3 subset it scores 49/90 against R4's 46/90.