Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
name: string
version: string
seeds: list<item: int64>
  child 0, item: int64
note: string
instances: list<item: struct<id: string, config: string, category: string, task: string, variant: string, knobs (... 60 chars omitted)
  child 0, item: struct<id: string, config: string, category: string, task: string, variant: string, knobs: struct<cu (... 48 chars omitted)
      child 0, id: string
      child 1, config: string
      child 2, category: string
      child 3, task: string
      child 4, variant: string
      child 5, knobs: struct<cue: string>
          child 0, cue: string
      child 6, seed: int64
      child 7, manifestSha256: string
to
{'instances': List({'id': Value('string'), 'task': Value('string'), 'seed': Value('int64'), 'cue': Value('string'), 'reflect': Value('bool')})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1835, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 621, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              name: string
              version: string
              seeds: list<item: int64>
                child 0, item: int64
              note: string
              instances: list<item: struct<id: string, config: string, category: string, task: string, variant: string, knobs (... 60 chars omitted)
                child 0, item: struct<id: string, config: string, category: string, task: string, variant: string, knobs: struct<cu (... 48 chars omitted)
                    child 0, id: string
                    child 1, config: string
                    child 2, category: string
                    child 3, task: string
                    child 4, variant: string
                    child 5, knobs: struct<cue: string>
                        child 0, cue: string
                    child 6, seed: int64
                    child 7, manifestSha256: string
              to
              {'instances': List({'id': Value('string'), 'task': Value('string'), 'seed': Value('int64'), 'cue': Value('string'), 'reflect': Value('bool')})}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1379, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 956, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1702, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1887, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

instances
list
[{"id":"base/backmost_layer_select/backmost_layer_select_s1","task":"backmost_layer_select","seed":1(...TRUNCATED)

TuringSpace-Gym

TuringSpace-Gym makes training trajectories on the 125 tasks of TuringSpace-Bench. A reference agent plays each instance with the mouse, one action at a time, and every action is saved with the screenshot it was chosen on; export_sft.py turns the episodes into the ms-swift samples TuringSpace-9B is trained on.

The tasks are in five categories of 25, one folder each under tasks/: A spatial relation, B closed-loop manipulation, C multi-step planning, D mental transformation, E spatiotemporal reasoning.

The benchmark, the evaluated agents and the training of TuringSpace-9B are in os-cua-framework/TuringSpace. seeds/train_instances.json lists 55,000 training instances: 500 per task in categories A–D and 200 per task in category E. model/ holds TuringSpace-9B, trained on the trajectories of these instances.

πŸ“ Contents

TuringSpace-Gym/
β”œβ”€β”€ tasks/              the task pages, the reference agent's gesture planner and the asset pools
β”œβ”€β”€ turingspace_gym/    the environment and the reference agent
β”œβ”€β”€ serve.py            static server of tasks/
β”œβ”€β”€ pick_seeds.py       picks training instances
β”œβ”€β”€ rollout.py          plays a list of instances
β”œβ”€β”€ export_sft.py       episodes -> ms-swift samples
β”œβ”€β”€ seeds/              train_instances.json, test_instances.json
β”œβ”€β”€ model/              TuringSpace-9B
└── docs/

πŸ€– Model

model/ is TuringSpace-9B: Qwen3.5-9B with full-parameter SFT on the 55,000 trajectories of seeds/train_instances.json (1,000 steps, global batch 128, learning rate 5e-6 with cosine decay to 1e-6, vision encoder frozen), a Hugging Face safetensors checkpoint in bfloat16. Its MTP layers are not included, so serve it without MTP speculative decoding.

hf download zhaihaotian/TuringSpace-Gym --repo-type dataset --include "model/*" --local-dir TuringSpace-Gym
vllm serve TuringSpace-Gym/model

βš™οΈ Setup

conda create -n turingspace-gym python=3.11 -y && conda activate turingspace-gym
pip install -r requirements.txt
python -m playwright install chromium

πŸš€ Quick start

python serve.py tasks 7870 &
python rollout.py seeds/train_instances.json runs/train --num_envs 16
python export_sft.py seeds/train_instances.json runs/train train.jsonl

On one 12-core machine, 12 workers play about 6,700 episodes an hour on the 100 tasks of categories A–D.

πŸ“š Docs

πŸ“„ License

The code is released under the Apache License 2.0 (LICENSE). The assets keep the licences of their sources, listed in tasks/assets/ATTRIBUTION.md. Do not use TuringSpace to bypass CAPTCHAs that protect real services, and do not train on the test instances (seeds 36, 43 and 50) of TuringSpace-Bench.

πŸ“ Citation

@article{ding2026turingspace,
  title   = {TuringSpace: Grounding Spatial Intelligence in Action with Interactive CAPTCHAs},
  author  = {Ding, Xuwei and Zhai, Skylar and Song, Linxin and Feng, Jiarui and Fu, Dongqi and Xi, Huajun and
             Li, Junnan and Liang, Jingcheng and Zhao, Jieyu and Xia, Yinglong and Li, Hong and Kang, Jian},
  journal = {arXiv preprint},
  year    = {2026}
}
Downloads last month
63