Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
image
label
class label
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
622ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3
7635253b65-1c19-4304-8aa4-6884b8218fc0
7635253b65-1c19-4304-8aa4-6884b8218fc0
7635253b65-1c19-4304-8aa4-6884b8218fc0
7635253b65-1c19-4304-8aa4-6884b8218fc0
7635253b65-1c19-4304-8aa4-6884b8218fc0
7635253b65-1c19-4304-8aa4-6884b8218fc0
7635253b65-1c19-4304-8aa4-6884b8218fc0
7635253b65-1c19-4304-8aa4-6884b8218fc0
1596c4c23a1-42a4-43cc-9db1-2f86ff3738cc
1596c4c23a1-42a4-43cc-9db1-2f86ff3738cc
1596c4c23a1-42a4-43cc-9db1-2f86ff3738cc
1596c4c23a1-42a4-43cc-9db1-2f86ff3738cc
1596c4c23a1-42a4-43cc-9db1-2f86ff3738cc
1937f52cab9-535c-4835-ac8c-391ee64dc930
1937f52cab9-535c-4835-ac8c-391ee64dc930
1937f52cab9-535c-4835-ac8c-391ee64dc930
1937f52cab9-535c-4835-ac8c-391ee64dc930
1937f52cab9-535c-4835-ac8c-391ee64dc930
1937f52cab9-535c-4835-ac8c-391ee64dc930
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
19782bc8d6a-36eb-4d2d-8801-ef714fb1e55a
End of preview. Expand in Data Studio

OSWorld1.0 UI-TARS-1.5-7B Baseline

This dataset contains the complete local baseline run of ByteDance-Seed/UI-TARS-1.5-7B on the 361-task OSWorld 1.0 test_nogdrive.json task set.

Result

  • Tasks completed: 361/361
  • Total OSWorld score: 55.8522061570
  • Average OSWorld score: 15.47%
  • Exact average: 0.15471525251257395
  • Tasks with score 1.0: 54
  • Tasks with a non-zero partial score: 3

The score above is the average of all 361 per-task result.txt values. It is the result of this reproduction run, not the score reported by the model authors.

Model and benchmark versions

  • Model: ByteDance-Seed/UI-TARS-1.5-7B
  • Model revision: 683d002dd99d8f95104d31e70391a39348857f4e
  • OSWorld repository commit: 315a7603173feadf1b8a85cbc006c93ffe1dc1a1
  • OSWorld task set: evaluation_examples/test_nogdrive.json
  • Official runner: scripts/python/run_multienv_uitars15_v1.py
  • Official agent: mm_agents/uitars15_v1.py

Evaluation configuration

  • Action space: pyautogui
  • Observation: screenshot only
  • Screen size: 1920 × 1080
  • Maximum steps per task: 100
  • Sleep after each action: 3.0 seconds
  • Model type: qwen25vl
  • Inference and prompt mode: qwen25vl_normal
  • History length: 5
  • Prompt language setting: Chinese
  • Input swap: enabled by the official runtime configuration
  • Temperature: 0.0
  • Top-p: 0.9
  • Top-k: -1
  • Maximum generated tokens: 1000
  • Maximum image pixels: 16384 × 28 × 28
  • Minimum image pixels: 100 × 28 × 28
  • Call-user tolerance: 3

Inference deployment

The model was served with the official TGI settings:

  • Text Generation Inference: 3.2.1
  • No quantization
  • Max input length: 65536
  • Max batch prefill tokens: 65536
  • Max total tokens: 65537
  • CUDA_GRAPHS=0
  • Payload limit: 8000000 bytes

The run used eight NVIDIA RTX 3090 GPUs. There were four independent evaluation shards; each shard had one TGI endpoint using two GPUs and one local OSWorld Docker environment.

The official TGI 3.2.1 container image was executed as a Singularity SIF because the machine's Docker data partition did not have room for another image. Model weights, TGI code, generation settings, prompts, action parser, and evaluator were unchanged.

Local OSWorld environment

  • Provider: local docker
  • One environment per shard
  • KVM hardware acceleration
  • OSWorld Ubuntu QCOW2 image
  • Four task shards: 91 / 90 / 90 / 90

The runner imports the AWS provider even in Docker mode. Dummy AWS environment values were supplied only to satisfy that import; no AWS instance was used.

The upstream generic OSWorld runner passes an unused vm_ip keyword to UITARSAgent.reset, while the upstream UI-TARS agent does not accept it. A minimal runtime compatibility shim ignored only this unused keyword. It did not change model inference, agent state, prompts, actions, or evaluation.

Some proxy-dependent web tasks encountered site connectivity failures, and some generated clipboard actions encountered PyperclipException in the local desktop container. These environment limitations are preserved in the logs and may have reduced the measured score.

Dataset layout

config/
  shard0.json ... shard3.json
logs/
  runner-shard*.log
  tgi-shard*.log
shard0/ ... shard3/
  pyautogui/screenshot/uitars15-7b/<domain>/<task-id>/
    traj.jsonl
    result.txt
    runtime.log
    recording.mp4
    step_*.png

Every task has a traj.jsonl and a result.txt. The trajectories include the model response, parsed action, reward, and environment state recorded by the official OSWorld runner. Step screenshots and recordings are included when produced by the runner.

Downloads last month
2,558