sokoban
Qwen3-30B-A3B-Thinking-2507-ECHO-Sokoban-GRPO-GGUFSPA-sokoban-qwen2.5-1.5b-instruct-GGUFVPR-Qwen3-4B-Sokoban1.5b-base-sokoban-sfted-wza-GGUFsokoban-3b-sfted-GGUFagentic-nonNorm-sokoban-NonMarkov-qwen2.5-3B_6x6_2-4_SFT_gm0_advGm0.5_actoragentic-nonNorm-sokoban-NonMarkov-qwen3-4B_6x6_2-4_SFT_gm0_advGm0.5_actoragentic-sokoban-NonMarkov-qwen3-4B_SFT-5e-6-4k_prm0_actor
Datasets
All datasets matching “sokoban”sokoban-diversity-trajectories
Sokoban Diversity Trajectories
Model reasoning-trajectory solution data for Sokoban
puzzles. Each record is a model's full chain-of-thought attempt at solving a puzzle, paired with a
machine-extracted move trace that has been replay-verified against the board. Only accepted
(puzzle-solving) trajectories are kept.
The dataset is designed to study how the diversity of the generator (single model vs. a fleet of
models vs. token-level cross-model continuation) affects the resulting… See the full description on the dataset page: https://huggingface.co/datasets/shizhuo2/sokoban-diversity-trajectories.sokoban-10k-vjepa2-tokenizedsokoban-solving
Sokoban solving levels
Procedurally generated Sokoban levels, solvable by construction: a random room is carved, the
boxes are placed on the targets, and the player pulls them away along a long random walk, so
replaying the walk backwards solves the level. Six tiers, from 12 by 12 with 6 boxes to 24 by 24
with 18 boxes.
data/heldout/s12b06.txt ... s24b18.txt 200 levels per tier, 1200 in all, no solutions published
data/train/s12b06.txt ... s24b18.txt 500 levels per tier… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/sokoban-solving.sokoban_easy_v8_res256_stopreq_allstep_cot_ordered100k
Sokoban Easy V8 Res256 Stop-Required All-Step CoT Ordered 100k
This repository contains the exact Sokoban easy all-step CoT ordered dataset used for the 2026-06-04 ordered-strict SFT run.
Archive parts:
sokoban_easy_v8_res256_stopreq_allstep_cot_ordered100k_20260604.tar.gz.part-*
Reassemble before extraction:
cat sokoban_easy_v8_res256_stopreq_allstep_cot_ordered100k_20260604.tar.gz.part-* > sokoban_easy_v8_res256_stopreq_allstep_cot_ordered100k_20260604.tar.gz
sha256sum -c… See the full description on the dataset page: https://huggingface.co/datasets/novastar113/sokoban_easy_v8_res256_stopreq_allstep_cot_ordered100k.sokoban-boxoban
Boxoban Sokoban levels
The Boxoban level sets released by DeepMind (Guez, Mirza, Gregor, Kabra et al., "An investigation
of model-free planning", 2018; https://github.com/google-deepmind/boxoban-levels), redistributed
unchanged under their Apache 2.0 licence, with the hard set split into two folders.
boxoban-levels.tar.gz the same files as one archive, for a build that fetches a single file
levels/unfiltered/{train,valid,test}/ 900000, 100000 and 1000 levels… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/sokoban-boxoban.sokobanlevels