SlowGuess/qwen38-minecraft-m2-data

Minecraft per-tick action sequences for VLM action SFT

This dataset contains frame + low-level action sequences for supervised fine-tuning of a vision-language model to play Minecraft with native keyboard and mouse actions, one action per 50 ms tick.

It is built from two public sources:

• the CraftJarvis OpenHA datasets (text-action, motion-coa, grounding-coa);
• the OpenAI VPT contractor recordings they come from.

What differs from the OpenHA release

Step	Directory	Change
Repaired	m2_repaired/	OpenHA's frame windows and task text are kept. Actions are recomputed from the original VPT .jsonl lines, because OpenHA rounds camera values toward zero to whole degrees. The VPT cursor is drawn on frames where a GUI is open; the OpenHA frames have no cursor. Duplicate labels of the same frame with the same task text are merged. Train/valid are split by recording (5 % valid, seed 42).
One wording per frame	m2_onepf/	When a frame is supervised under several task paraphrases, one paraphrase is kept, in priority text-action > motion-coa > grounding-coa.
Recipe-book sessions	vpt_book/	Human crafting sessions in which the recipe book is clicked, cut from VPT recordings (cursor drawn). Sessions with typing in the search box are removed, as are M2 validation recordings and frames already supervised in m2_onepf.
Merged	m2_onepf_book/	m2_onepf train + vpt_book train.
200 ms chunks	chunk200/	m2_onepf_book train and m2_repaired valid in the 200 ms action-chunk format of Lumine (arXiv 2511.08892) and Scaling in Games (arXiv 2610.28698). Frames are stitched back into continuous segments of one event. Each segment is cut into chunks of 4 ticks, and the leftover ticks at a segment end are dropped. Windows hold 20 chunks with a 10-chunk stride. Used to train chunk200/ in SlowGuess/qwen38-minecraft-r4.
Ablation E1	chunk200_e1/	Each m2_onepf_book row (and m2_repaired valid row) on its own, in the 200 ms format. There is no stitching and no overlap. Chunks are aligned to the row end; leftover ticks at the row start are dropped, and rows shorter than 4 ticks are dropped.
Ablation E2	chunk200_e2/	Same as chunk200/, but with a 20-chunk stride, so windows do not overlap and every chunk is supervised once.

Contents

Path	Rows (sequences)	Supervised actions	Size
m2_repaired/seq/train.jsonl	332,499	8,198,765	2.10 GB
m2_repaired/seq/valid.jsonl	17,558	431,527	0.11 GB
m2_onepf/seq/train.jsonl	239,951	5,690,654	1.46 GB
vpt_book/seq/train.jsonl	18,670	478,677	0.13 GB
m2_onepf_book/seq/train.jsonl	258,621	6,169,331	1.59 GB
chunk200/seq/train.jsonl	151,285	2,678,233 chunks (1,523,284 distinct)	0.82 GB
chunk200/seq/valid.jsonl	7,423	128,398 chunks	0.04 GB
chunk200_e1/seq/train.jsonl	217,381	1,427,468 chunks	0.46 GB
chunk200_e1/seq/valid.jsonl	15,236	99,864 chunks	0.03 GB
chunk200_e2/seq/train.jsonl	101,331	1,523,284 chunks	0.47 GB
chunk200_e2/seq/valid.jsonl	4,999	73,200 chunks	0.02 GB
openha/images_zip/	16,549 zip files (one per recording; 3,238 under samples_v6/)		183.5 GB
m2_repaired/images_zip/	5,994 zip files (1,166,464 new or cursor-drawn frames)		38.6 GB
vpt_book/images_zip/	2,363 zip files (549,487 frames)		23.4 GB
m2_repaired/vpt_actions.jsonl.gz	per-tick actions and cursor positions read from VPT .jsonl for 13,302 recordings		189 MB
m2_repaired/recordings.txt, vpt_book/segments_book.json, vpt_book/event_description.json, */seq/summary.json	recording list, recipe-book segments, OpenHA task wordings, build statistics		small

• All validation is m2_repaired/seq/valid.jsonl. On the build machine, m2_onepf/seq/valid.jsonl and m2_onepf_book/seq/valid.jsonl are symlinks to it.
• Each sequence holds 1–30 consecutive ticks from one recording and is at most 16,384 tokens with the Qwen3.8 tokenizer.

Row format

{"id": "text-action:0:0", "source_dataset": "text-action", "source_row_id": "...-1403",
 "source_recording": "/scratch/yz979_yale/openha/data/openha/images/images/<recording>",
 "task": "Mine the cobweb to clear the path.",
 "steps": [{"image": ".../images/images/<recording>/<recording>-1403.jpg",
            "action": {"camera": [-0.6, -8.55], "left": 1}, "source_index": 0, "supervise": true}, ...]}

• camera is [pitch, yaw] in degrees. Buttons and keys that are pressed appear with value 1.
• An empty action means no input on that tick.

200 ms chunk rows (chunk200*/)

{"id": "chunk200:<recording>:1074", "source_recording": "<recording>", "event": "mine_block:tall_grass",
 "task": "Collect tall grass by breaking it.",
 "steps": [{"image": ".../images/images/<recording>/<recording>-1074.jpg",
            "action": {"keys": [["forward"], ["forward"], ["forward"], ["forward"]]}, "source_index": 0, "supervise": true},
           {"image": ".../<recording>-1078.jpg",
            "action": {"camera": [1.5, 5.25], "keys": [["forward"], ["forward"], ["forward"], ["attack", "forward"]]},
            "source_index": 1, "supervise": true}, ...]}

• Each step is one chunk. image is the chunk's first frame.
• keys lists the keys held on each of the 4 ticks.
• camera is the sum of the 4 ticks' camera values, which the policy executes evenly over the 4 ticks. It is omitted when the sum is zero.
• chunk200_e1 rows carry R4's row id (e1:<id>) and its absolute source_recording path, and have no event field.
• seq/summary.json in each folder gives the build statistics and the exact cutting rules.
• The build scripts are prepare_qwen38_chunk_sequences.py (chunk200, and chunk200_e2 with stride 20) and prepare_qwen38_chunk_rows.py (chunk200_e1).

Image paths

• image fields are absolute paths on the machine that built the data, under three prefixes:

Prefix in image	Directory in this repo
/scratch/yz979_yale/openha/data/openha/	openha/
/work/yale/ac3458_ea_yale/openha/data/m2_repaired/	m2_repaired/
/work/yale/ac3458_ea_yale/openha/data/vpt_book/	vpt_book/

• A frame <prefix>/images/images/<recording>/<file>.jpg is the member <file>.jpg of <prefix>/images_zip/<recording>.zip. samples_v6 recordings live one level deeper: images_zip/samples_v6/<recording>.zip.
• Hugging Face allows at most 10k files per folder. The 13,311 top-level OpenHA zips are therefore stored under one-character subfolders by the first letter of the recording name. Put them back after downloading:

cd openha/images_zip && for d in ?; do mv "$d"/*.zip . && rmdir "$d"; done

• To use the .jsonl files unchanged, symlink the three prefixes to the downloaded directories, or rewrite the prefixes in the .jsonl files.
• The chunk200*/ files use the same frames and the same prefixes.

Sources and license

• The OpenHA datasets are released by CraftJarvis under the MIT license.
• The VPT recordings come from OpenAI's Video-Pre-Training repository (MIT).
• Minecraft is the intellectual property of Microsoft.
• This derived dataset is released under MIT. Please also cite OpenHA (arXiv 2509.13347) and VPT (arXiv 2206.11795).