RynnUnify-GEU-Bench

RynnUnify-GEU-Bench is the GEU (Generation Enhanced Understanding) benchmark used for the RynnUnify family. It contains 2000 multi-turn samples across 10 subsets and 3 task families, each paired with its source images and a stored rollout video.

Each record walks a model through the same three-stage protocol:

1. World modeling — given the original image(s) or video and the question, emit <world_modeling>…</world_modeling> and <duration>…</duration>.
2. Rollout observation — the stored rollout video referenced by the record is supplied as a new user turn.
3. Answer — emit <thinking>…</thinking> and <answer>…</answer>.

Contents

Property	Value
Records	2000
Subsets (200 each)	position, direction, distance, size, physics, mess, affordance, area, action reasoning, trajectory
Image files	2571
Video files	2200
Media size	2,237,117,321 bytes (2.08 GiB)

Task family (task_level_1)	Subsets	Records
spatial consolidation	position, direction, distance, size	800
action rehearsal	physics, mess, affordance, area	800
dynamic rollout	action reasoning, trajectory	400

RynnUnify-GEU-Bench/
├── rynnunify_geu_bench.jsonl
├── cases/
│   └── <subset>/<case>/          e.g. cases/position/0000/
│       ├── image_0.JPG           original image(s) for the first user turn
│       ├── image_1.JPG           (image_1 is present only for two-image cases)
│       └── video_0.mp4           stored rollout video for the second user turn
├── eval/
│   ├── README.md                 (protocol and scoring reference)
│   ├── eval_rynnomni_api.py      (shared runner and scoring)
│   ├── eval_rynnomni_qwen3.py    (local Qwen3-VL inference)
│   ├── eval_rynnomni_qwen25.py   (local Qwen2.5-VL inference)
│   ├── eval_rynnomni_cosmos3.py  (Cosmos3 inference)
│   ├── geu_reference_adapter.py  (two-turn oracle rollout reference)
│   └── inference_geu.py          (single-GPU reference inference)
└── README.md

<subset> is the lowercased, underscore-separated task_level_2: position, direction, distance, size, physics, mess, affordance, area, action_reasoning, trajectory.

Record schema

{
  "task_level_1": "spatial consolidation",
  "task_level_2": "position",
  "conversation": [
    {"role": "system", "content": [{"type": "text", "text": "…"}]},
    {"role": "user", "content": [
      {"type": "image", "image": "cases/position/0000/image_0.JPG"},
      {"type": "image", "image": "cases/position/0000/image_1.JPG"},
      {"type": "text", "text": "Looking at image 1 and image 2 …"}
    ]},
    {"role": "assistant", "content": [{"type": "text", "text": "<world_modeling>…</world_modeling><duration>4s</duration>"}]},
    {"role": "user", "content": [
      {"type": "video", "video": "cases/position/0000/video_0.mp4"},
      {"type": "text", "text": "The following is the stored rollout …"}
    ]},
    {"role": "assistant", "content": [{"type": "text", "text": "<thinking>…</thinking><answer>…</answer>"}]}
  ]
}

Media paths in conversation[*].content[*].image|video are relative to this repository root (the directory containing rynnunify_geu_bench.jsonl), so no path remapping is needed when the dataset is downloaded in full. Absolute mount prefixes can still be replaced with --path_map OLD=NEW. Every case directory is self-contained: <case> is the 0-based record index, images are numbered in turn order, and the rollout video is always video_0.mp4 (a few first-turn video cases store both the input video as video_0.mp4 and the rollout as video_1.mp4; follow the JSONL paths).

Evaluation

The bundled harness in eval/ is the reference implementation (inference_geu.py oracle-rollout protocol plus 8-GPU Qwen/Cosmos runners and the GPT-based scorer for mess/affordance QA):

pip install 'transformers>=4.57.0' accelerate torchvision av pillow openai 'httpx[http2]' tqdm

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python eval/eval_rynnomni_qwen3.py \
  --data ./rynnunify_geu_bench.jsonl \
  --model_path /path/to/Qwen3-VL-checkpoint \
  --output_dir ./results/qwen3_8gpu --num_gpus 8 --dtype bf16 \
  --mode oracle_rollout --fps 2 --max-frames 512

Add --inference_only to skip GPT scoring, and --eval_only to rescore cached predictions. See eval/README.md for per-subset prompting rules, scoring semantics (0–100 partial credit for mess, continuous relative error for distance) and resume behaviour.

Provenance and license

The records and media were assembled from the internal GEU collections (OMNI_DATA/GEU_DATA, scene-understanding frames, and stored rollout videos) and normalized so that every reference resolves inside this package. Media derive from their upstream corpora and remain subject to those corpora's terms; they are released here for non-commercial research and evaluation. The evaluation code follows the reference implementation and is provided under the same research-use terms. Contact the RynnUnify authors for other uses.

Citation

@misc{rynnunify2026,
  title={RynnUnify: Towards Unified Embodied Brain},
  author={Minghao Zhu and Bohan Hou and Kun Xiang and Zixuan Wang and Yinzhou Tang and Yifei Yang and Zhikai Wang and Tianyi Zhang and Kehan Li and Haoyu Zhao and Siteng Huang and Leigang Qu and Sicong Leng and Dongchi Huang and Mingxiu Chen and Shijian Lu and Yue Wang and Xiaodan Liang and Zhenyu Zeng and Deli Zhao and Chengju Liu and Jianfei Yang and Xin Li},
  year={2026},
}