| --- |
| license: mit |
| task_categories: |
| - robotics |
| - video-text-to-text |
| tags: |
| - minecraft |
| - vla |
| - vision-language-action |
| - instruction-following |
| size_categories: |
| - 100K<n<1M |
| --- |
| |
| # Minecraft VLA Stage 2: Instruction-Following Data |
|
|
| Stage 2 of the TESS-Minecraft Vision-Language-Action training pipeline. |
|
|
| ## Overview |
|
|
| This dataset adds **task instructions** to the Stage 1 visuomotor data, enabling instruction-following training. |
|
|
| ## Data Format |
|
|
| | Field | Type | Description | |
| |-------|------|-------------| |
| | `id` | string | Unique sample ID | |
| | `video_id` | string | Source video name | |
| | `frame_idx` | int | Frame index within video | |
| | `instruction` | string | Task instruction (empty for continuation frames) | |
| | `action` | string | Action in VLA format | |
| | `task_category` | string | Task type (e.g., "mine_block:oak_log") | |
| | `task_group` | string | High-level task (e.g., "mine_block") | |
| | `target` | string | Target object (e.g., "oak_log") | |
| | `subset` | string | VPT subset (6.x, 7.x, 10.x) | |
| | `is_segment_start` | bool | True for first frame of each task segment | |
| | `image_bytes` | bytes | JPEG-encoded frame (640x360) | |
|
|
| ## Methodology |
|
|
| **Source:** Matches [Stage 1 data](https://huggingface.co/datasets/TESS-Computer/minecraft-vla-stage1-clean) with [JARVIS-VLA](https://huggingface.co/datasets/CraftJarvis/minecraft-vla-sft) instructions. |
|
|
| **Qwen3-Style Training Format:** Instructions are provided ONLY at segment start (`is_segment_start=True`). Continuation frames have empty instruction - the model must maintain goal context. This matches how the model is used at inference: give instruction once, then stream images. |
|
|
| ## Usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("TESS-Computer/minecraft-vla-stage2-clean", split="train", streaming=True) |
| |
| for sample in ds: |
| instruction = sample['instruction'] |
| action = sample['action'] |
| image = sample['image_bytes'] # JPEG bytes |
| break |
| ``` |
|
|
| ## Links |
|
|
| - [Stage 1 Data](https://huggingface.co/datasets/TESS-Computer/minecraft-vla-stage1-clean) |
| - [Project Repository](https://github.com/TESS-Computer/TESS-Minecraft) |
| - [JARVIS-VLA](https://craftjarvis.github.io/JarvisVLA/) |
| - [Lumine Paper](https://www.lumine-ai.org/) |
|
|
