DarthVaderSenior/visgym-pacman-2d

---
license: other
task_categories:
- image-to-text
- reinforcement-learning
pretty_name: VisGym Pacman2D Deterministic Trajectories
---

# VisGym Pacman2D Deterministic Trajectories

This dataset contains behavior-cloning trajectories for the custom VisGym Pacman2D environment.

Each row is one successful oracle episode. The history entries include `image_prev`, `image`, and
`image_next`; no synthetic thinking traces are stored. Images are JPEG base64 strings rendered from
the greyblue9/pacman-python visual assets used by the environment.

Environment summary:

- Grid size is constrained to 7x7 through 12x12.
- Easy uses the 9x9 setting.
- Hard uses the 11x11 setting with 8 food pellets.
- Ultrahard uses the 11x11 hard layout with 8 food pellets and a 40-step budget.
- Entities are Pacman, sparse food, walls, and exactly one deterministic ghost.
- The ghost chases Pacman by BFS distance with fixed tie-break order up, down, left, right.
- The movement action space is four directions; VisGym records can also end with `('stop', 'stop')`.

Splits:

- `data/pacman_2d_easy_v0/train/*.jsonl.gz`
- `data/pacman_2d_easy_v0/test/*.jsonl.gz`
- `data/pacman_2d_hard_v0/train/*.jsonl.gz`
- `data/pacman_2d_hard_v0/test/*.jsonl.gz`
- `data/pacman_2d_ultrahard_v0/train/*.jsonl.gz`
- `data/pacman_2d_ultrahard_v0/test/*.jsonl.gz`

Train/test initial states are checked to be disjoint by `init_state_hash`; see
`metadata/init_state_hash_audit.json`.

Target Hub repo: `https://huggingface.co/datasets/DarthVaderSenior/visgym-pacman-2d`