novastar111/sokoban_medium

---
pretty_name: Sokoban Medium
task_categories:
- reinforcement-learning
- other
tags:
- sokoban
- planning
- synthetic-data
- visgym
size_categories:
- 100K<n<1M
configs:
- config_name: default
data_files:
- split: test
path: trajectories/sokoban_medium/test/*.jsonl
- split: train
path: trajectories/sokoban_medium/train/*.jsonl
---

# sokoban_medium

Synthetic Sokoban dataset generated from the local VisGym Sokoban environment.

## Contents

- `trajectories/sokoban_medium/test/*.jsonl`
- `trajectories/sokoban_medium/train/*.jsonl`
- `manifests/`
- `metadata/`

## Generation Summary

- Task: `sokoban_medium`
- Raw test generated: `1000`
- Final test after dedupe: `1000`
- Raw train generated: `200000`
- Final train after dedupe: `199437`
- Train samples removed by test-hash filter: `500`
- Train samples removed by train self-dedupe: `63`
- Raw train chunk count: `20`

## Leakage / Hash Logic

- Hash basis: `init_state` only
- Test split:
- raw generation first
- dedupe within test
- write `metadata/test_init_state_hashes.json`
- Train split:
- raw generation only during chunked production
- final dedupe order:
1. remove records whose `init_state` hash appears in test
2. remove duplicate `init_state` hashes within train
- Final leakage report:
- `clean = True`
- `overlap_total = 0`

## Notes

- This dataset intentionally keeps the visual assets and prompt/action semantics from the integrated Sokoban environment while using the repo's current generation pipeline.
- Review artifacts are not included in the upload path by default.