SPARE GPT-5.5 Grounded Cognitive Multi-Turn Games

This public dataset contains 7,872 validated Python game environments for actor-only SPARE training.

• Six cognitive skills, exactly 1,312 environments per skill
• Generated with GPT-5.5 and grounded by spice_megascience_15k.jsonl
• Grounding corpus SHA-256: a36a928b4940b5b5d9e3f4cb5804a94c69462360943adb3be14613c82f0f72c0
• Maximum 25 turns and 32K generation context
• Every environment passes load, reset, step, and replay validation with a 2-second reset budget
• No tool-use or terminal environments

Layout

• games/*.py: trainer-ready SPARE environment files
• data/train.jsonl: Hugging Face datasets representation with embedded source code
• manifest.jsonl: generation, skill, grounding, source, and checksum metadata
• validation_manifest.jsonl: per-environment validation results
• validation_summary.json: validation totals and parameters
• checksums.sha256: SHA-256 checksums for every game file

Training

The pool supports 400 rollout steps with 24 environments per step. SPARE's static no-replacement sampler exhausts and reshuffles the full pool before reuse, so 7,872 environments are seen once before 1,728 controlled repeats.

hf download simonycl/spare-gpt55-cognitive-multiturn-7872 \
  --repo-type dataset \
  --local-dir /scratch/spare-gpt55-cognitive-multiturn-7872

Point STATIC_POOL_DIR at the downloaded games/ directory.

Citation

@article{liu2026spade,
  title={SPADE: Self-Play in Adaptive Synthetic Executable Environments},
  author={Liu, Bo and Yu, Simon and Jiang, Yiding and Qu, Ao and Zhao, Andrew and Liu, Zichen and Kim, Junsu and Zhou, Zijian and Kim, Seungone and Ren, Tongzheng and Liu, Mickel and Yu, Hanfei and Chen, Zhaorun and Shi, Weiyan and Liang, Paul Pu and Zettlemoyer, Luke and Choi, Yejin and Jaques, Natasha},
  journal={arXiv preprint arXiv:2608.19197},
  year={2026}
}