datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
sokobanlevelssokoban_easy_cot_chunk_k1_train
sokoban_easy_cot_chunk_k1_train
BAGEL VLM-Gym world-model dataset (sokoban / cot).
CoT chunk-K train set: all-step interleaved imagined reasoning; re-grounds on the true frame every K=1 steps.
layout: Train-only. Gzipped-JSONL shards under training/; each row is one packed SFT sample with base64-JPEG frames inline.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org; CoT and non-CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_easy_cot_chunk_k1_train.sokoban_hard_3box_remap_prev
VisGym sokoban_hard_3box Remap-Prev
Solver-success Sokoban trajectories generated from the VisGym Sokoban environment.
HF repo: https://huggingface.co/datasets/novastar112/sokoban_hard_3box_remap_prev
task: sokoban_hard_3box
env_id: sokoban/hard
env_kwargs: {"action_mode": "move_only", "dim_room": [10, 10], "max_steps": 40, "num_boxes": 3, "num_gen_steps": 34, "success_mode": "stop_then_solved"}
action schema: move-only ('move', 1..4) plus ('stop', 'stop')
success semantics:… See the full description on the dataset page: https://huggingface.co/datasets/novastar112/sokoban_hard_3box_remap_prev.Sokoban-Trajectories
Sokoban-Trajectories
Rollout trajectories for the Sokoban (推箱子) environment generated using the
RAGEN framework.
Models
Role
Model
Teacher
Qwen/Qwen2.5-14B-Instruct
Student
Qwen/Qwen2.5-3B-Instruct
Environment Settings
Setting
Value
Room size
6×6
Number of boxes
1
Max turns per episode
10
Actions per turn
up to 2 (max_actions_per_turn=2)
Max actions per trajectory
20 (10 turns × 2 actions)
Note: 1 turn = up to 2… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/Sokoban-Trajectories.sokoban_easy_1box_remap_prev
VisGym sokoban_easy_1box Remap-Prev
Solver-success Sokoban trajectories generated from the VisGym Sokoban environment.
HF repo: https://huggingface.co/datasets/novastar112/sokoban_easy_1box_remap_prev
task: sokoban_easy_1box
env_id: sokoban/easy
env_kwargs: {"action_mode": "move_only", "dim_room": [7, 7], "max_steps": 20, "num_boxes": 1, "num_gen_steps": 23, "success_mode": "stop_then_solved"}
action schema: move-only ('move', 1..4) plus ('stop', 'stop')
success semantics:… See the full description on the dataset page: https://huggingface.co/datasets/novastar112/sokoban_easy_1box_remap_prev.sokoban_adaptive_rl_2box_train_2000
sokoban_adaptive_rl_2box_train_2000
BAGEL VLM-Gym world-model dataset (sokoban / rl_train).
Fresh mixed 2-box RL training pool, disjoint from SFT data and the frozen evaluation suite.
layout: Held-out eval episode pool. shard_*.jsonl.gz at the repo root; one episode per row with a contiguous global index.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org; CoT and non-CoT
variants… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_adaptive_rl_2box_train_2000.verl_agent_sokoban_dapo_qwen3_1.7b_noentropy_rolloutsokoban_easy_v8_noncot_chunk_k10_world_model_20260622_perseg
sokoban_easy_v8_noncot_chunk_k10_world_model_20260622_perseg
Sokoban action-conditioned visual world-model SFT data (non-CoT baseline) for the BAGEL-7B-MoT
VLM-Gym feedback-interval study.
Format: gzipped JSONL shards under training/, one packed row = one episode. Frames are
base64 JPEG (q95). Per-segment CoT layout: <think> per-step imagined frame (MSE target) </think>
then the committed action chunk; between chunks a loss-0 "Action executed." + real frame
(GT re-grounding).… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/sokoban_easy_v8_noncot_chunk_k10_world_model_20260622_perseg.sokoban_easy_v8_noncot_chunk_k5_world_model_20260622_perseg
sokoban_easy_v8_noncot_chunk_k5_world_model_20260622_perseg
Sokoban action-conditioned visual world-model SFT data (non-CoT baseline) for the BAGEL-7B-MoT
VLM-Gym feedback-interval study.
Format: gzipped JSONL shards under training/, one packed row = one episode. Frames are
base64 JPEG (q95). Per-segment CoT layout: <think> per-step imagined frame (MSE target) </think>
then the committed action chunk; between chunks a loss-0 "Action executed." + real frame
(GT re-grounding).… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/sokoban_easy_v8_noncot_chunk_k5_world_model_20260622_perseg.sokoban_easy_cot_chunk_kinf_train
sokoban_easy_cot_chunk_kinf_train
BAGEL VLM-Gym world-model dataset (sokoban / cot).
CoT chunk-K train set: all-step interleaved imagined reasoning; re-grounds on the true frame every K=inf (open-loop; imagine the whole episode) steps.
layout: Train-only. Gzipped-JSONL shards under training/; each row is one packed SFT sample with base64-JPEG frames inline.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s)… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_easy_cot_chunk_kinf_train.sokoban_easy_v8_noncot_chunk_k3_world_model_20260622_perseg
sokoban_easy_v8_noncot_chunk_k3_world_model_20260622_perseg
Sokoban action-conditioned visual world-model SFT data (non-CoT baseline) for the BAGEL-7B-MoT
VLM-Gym feedback-interval study.
Format: gzipped JSONL shards under training/, one packed row = one episode. Frames are
base64 JPEG (q95). Per-segment CoT layout: <think> per-step imagined frame (MSE target) </think>
then the committed action chunk; between chunks a loss-0 "Action executed." + real frame
(GT re-grounding).… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/sokoban_easy_v8_noncot_chunk_k3_world_model_20260622_perseg.sokoban_easy_noncot_chunk_k10_train
sokoban_easy_noncot_chunk_k10_train
BAGEL VLM-Gym world-model dataset (sokoban / noncot).
Non-CoT chunk-K train set (no imagined reasoning); re-grounds every K=10 steps.
layout: Train-only. Gzipped-JSONL shards under training/; each row is one packed SFT sample with base64-JPEG frames inline.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org; CoT and non-CoT
variants share the same… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_easy_noncot_chunk_k10_train.sokoban_easy_v8_cot_chunk_kinf_world_model_20260707_perseg
sokoban_easy_v8_cot_chunk_kinf_world_model_20260707_perseg
Sokoban action-conditioned visual world-model SFT data (CoT self-rollout) for the BAGEL-7B-MoT
VLM-Gym feedback-interval study.
Format: gzipped JSONL shards under training/, one packed row = one episode. Frames are
base64 JPEG (q95). Per-segment CoT layout: <think> per-step imagined frame (MSE target) </think>
then the committed action chunk; between chunks a loss-0 "Action executed." + real frame
(GT re-grounding). See… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/sokoban_easy_v8_cot_chunk_kinf_world_model_20260707_perseg.sokoban_easy_noncot_chunk_k1_train
sokoban_easy_noncot_chunk_k1_train
BAGEL VLM-Gym world-model dataset (sokoban / noncot).
Non-CoT chunk-K train set (no imagined reasoning); re-grounds every K=1 steps.
layout: Train-only. Gzipped-JSONL shards under training/; each row is one packed SFT sample with base64-JPEG frames inline.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org; CoT and non-CoT
variants share the same… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_easy_noncot_chunk_k1_train.sokoban_hard_allstep_thinking_cot_v8_res256_stop_prompt
Sokoban hard allstep CoT v8 res256 stop-prompt rebuild
This local dataset rebuilds the HF hard allstep-thinking CoT source against the v8 stop-required 256x256 state-replay hard train source. Rows removed by v8 layout dedup are skipped. Output shards are sorted to match the v8 hard train batch filenames and row order.
Summary: metadata/rebuild_summary.json
verl_agent_sokoban_grpo_qwen3_1.7b_subgoalver3_rolloutsokoban_adaptive_3box_deadlock_hard_250
sokoban_adaptive_3box_deadlock_hard_250
BAGEL VLM-Gym world-model dataset (sokoban / eval).
250 hard 3-box deadlock episodes across five temptation quantiles (corner + subtle); train-disjoint deadlock-capability benchmark.
layout: Held-out eval episode pool. shard_*.jsonl.gz at the repo root; one episode per row with a contiguous global index.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_adaptive_3box_deadlock_hard_250.sokoban_easy_eval_100_20260712
Sokoban ordinary easy — exact 100 episodes
Exact frozen Sokoban easy evaluation shard used by the UWM chunk-model trajectories.
Records: 100
Action mode: move-only
Success: stop-required and solved
Compressed JSONL SHA-256: 235754d970ad40b97f0469f7a9217922a1b6e4cd7c69735f9f0855b4a7b9ea41
Trajectories: https://huggingface.co/datasets/novastar111/uwm_paper_eval_trajectories
sokoban_easy_noncot_chunk_k3_train
sokoban_easy_noncot_chunk_k3_train
BAGEL VLM-Gym world-model dataset (sokoban / noncot).
Non-CoT chunk-K train set (no imagined reasoning); re-grounds every K=3 steps.
layout: Train-only. Gzipped-JSONL shards under training/; each row is one packed SFT sample with base64-JPEG frames inline.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org; CoT and non-CoT
variants share the same… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_easy_noncot_chunk_k3_train.sokoban_easy_v8_cot_chunk_k1_world_model_20260622_perseg
sokoban_easy_v8_cot_chunk_k1_world_model_20260622_perseg
Sokoban action-conditioned visual world-model SFT data (CoT self-rollout) for the BAGEL-7B-MoT
VLM-Gym feedback-interval study.
Format: gzipped JSONL shards under training/, one packed row = one episode. Frames are
base64 JPEG (q95). Per-segment CoT layout: <think> per-step imagined frame (MSE target) </think>
then the committed action chunk; between chunks a loss-0 "Action executed." + real frame
(GT re-grounding). See… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/sokoban_easy_v8_cot_chunk_k1_world_model_20260622_perseg.sokoban_easy_v8_cot_chunk_kinf_forward_20260710_fwdpure
sokoban_easy_v8_cot_chunk_kinf_forward_20260710_fwdpure
Sokoban action-conditioned visual world-model SFT data (CoT self-rollout) for the BAGEL-7B-MoT
VLM-Gym feedback-interval study.
Format: gzipped JSONL shards under training/, one packed row = one episode. Frames are
base64 JPEG (q95). Per-segment CoT layout: <think> per-step imagined frame (MSE target) </think>
then the committed action chunk; between chunks a loss-0 "Action executed." + real frame
(GT re-grounding). See… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/sokoban_easy_v8_cot_chunk_kinf_forward_20260710_fwdpure.sokoban_easy_cot_chunk_k3_train
sokoban_easy_cot_chunk_k3_train
BAGEL VLM-Gym world-model dataset (sokoban / cot).
CoT chunk-K train set: all-step interleaved imagined reasoning; re-grounds on the true frame every K=3 steps.
layout: Train-only. Gzipped-JSONL shards under training/; each row is one packed SFT sample with base64-JPEG frames inline.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org; CoT and non-CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_easy_cot_chunk_k3_train.sokoban_easy_noncot_chunk_k5_train
sokoban_easy_noncot_chunk_k5_train
BAGEL VLM-Gym world-model dataset (sokoban / noncot).
Non-CoT chunk-K train set (no imagined reasoning); re-grounds every K=5 steps.
layout: Train-only. Gzipped-JSONL shards under training/; each row is one packed SFT sample with base64-JPEG frames inline.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org; CoT and non-CoT
variants share the same… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_easy_noncot_chunk_k5_train.sokoban_adaptive_3box_hard_mixed_350
sokoban_adaptive_3box_hard_mixed_350
BAGEL VLM-Gym world-model dataset (sokoban / eval).
350 hard 3-box episodes: deadlock + trivial mixed in one contiguous pool; train-disjoint adaptive-thinking (when-to-think) benchmark.
layout: Held-out eval episode pool. shard_*.jsonl.gz at the repo root; one episode per row with a contiguous global index.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_adaptive_3box_hard_mixed_350.sokoban_compute_matched_3box_sft_train_2000
sokoban_compute_matched_3box_sft_train_2000
BAGEL VLM-Gym world-model dataset (sokoban / sft_control).
BFS-expert K=inf demonstrations built from the frozen 3-box RL boards for the compute-matched SFT control.
layout: Train-only. Gzipped-JSONL shards under training/; each row is one packed SFT sample with base64-JPEG frames inline.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org;… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_compute_matched_3box_sft_train_2000.sokoban_medium
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… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_medium.sokoban_v6_deadend
Sokoban V6 Dead-End Probe
This dataset is a 100-record Sokoban easy test probe enriched for early dead-end decisions.
Files
trajectories/sokoban_easy/test/batch_000000.jsonl: eval-compatible trajectory JSONL.
metadata/selection_summary.json: selection and BFS-validation summary.
metadata/leakage_report.json: train-split leakage hash audit.
metadata/leakage_overlaps.jsonl: all detected overlaps, including expected test-split source overlaps.… See the full description on the dataset page: https://huggingface.co/datasets/novastar112/sokoban_v6_deadend.sokoban_easy_cot_kinf_trickiness_train
sokoban_easy_cot_kinf_trickiness_train
BAGEL VLM-Gym world-model dataset (sokoban / trickiness).
Open-loop CoT train set with trickiness-graded reasoning (fires reasoning at graded-difficulty steps).
layout: Train-only. Gzipped-JSONL shards under training/; each row is one packed SFT sample with base64-JPEG frames inline.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org; CoT and… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_easy_cot_kinf_trickiness_train.sokoban_easy_mixed_deadlock_trivial
sokoban_easy_mixed_deadlock_trivial
BAGEL VLM-Gym world-model dataset (sokoban / eval).
Deadlock (temptation tiers) + trivial (no-deadlock) episodes in one contiguous pool; train-disjoint elicitation eval.
layout: Held-out eval episode pool. shard_*.jsonl.gz at the repo root; one episode per row with a contiguous global index.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org; CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_easy_mixed_deadlock_trivial.sokoban_adaptive_rl_3box_train_2000
sokoban_adaptive_rl_3box_train_2000
BAGEL VLM-Gym world-model dataset (sokoban / rl_train).
Mixed 3-box RL training pool: 1000 certified-deadlock and 1000 trivial boards.
layout: Held-out eval episode pool. shard_*.jsonl.gz at the repo root; one episode per row with a contiguous global index.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching sokoban checkpoint(s) under the companion model org; CoT and non-CoT
variants share the same… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/sokoban_adaptive_rl_3box_train_2000.