datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
pacman_hard_cot_chunk_k10_train
pacman_hard_cot_chunk_k10_train
BAGEL VLM-Gym world-model dataset (pacman / cot).
CoT chunk-K train set: all-step interleaved imagined reasoning; re-grounds on the true frame 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 pacman checkpoint(s) under the companion model org; CoT and non-CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_hard_cot_chunk_k10_train.pacman_v0
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… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_v0.pacman_hard_singleshot_train
pacman_hard_singleshot_train
BAGEL VLM-Gym world-model dataset (pacman / singleshot).
Single-shot open-loop train set (non-CoT; imagine the whole episode).
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 pacman checkpoint(s) under the companion model org; CoT and non-CoT
variants share the same underlying… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_hard_singleshot_train.pacman_braidthin_noncot_chunk_k5_train
pacman_braidthin_noncot_chunk_k5_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman 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/pacman_braidthin_noncot_chunk_k5_train.pacman_hard_cot_chunk_k3_train
pacman_hard_cot_chunk_k3_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman checkpoint(s) under the companion model org; CoT and non-CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_hard_cot_chunk_k3_train.pacman_eval_forward_stress_g12_random100_20260818
Forward stress G12 — deterministic random 100
This is the exact 100-episode Pacman evaluation shard used for the final UWM trajectories.
Source dataset: pacman_2d_easy_g12_f2_greedyfood_trap_forward_eval_v1_20260818
Sampling: Python MT19937 without replacement, seed 2026081812
Records: 100
JSONL SHA-256: 8118f7b275d345a1a6558df4f53c6de3382b23d8c663ebd112660e52219a6051
Trajectories: https://huggingface.co/datasets/novastar111/uwm_paper_eval_trajectories
sample_manifest.json… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_eval_forward_stress_g12_random100_20260818.pacman_2d_ultrahard_remap_imagined_rollout_5000
Pacman Ultrahard Source-Based Imagined Rollout Dataset
Source: novastar112/visgym_pacman_2d_remap
Task: pacman_2d_ultrahard_v0
Records: 5,000
Files:
trajectories/pacman_imagined_rollout.jsonl.gz: source trajectory records with one selected step containing four imagined one-step side rollouts.
transitions/pacman_imagined_rollout_transitions.jsonl.gz: flattened real and imagined transitions for world-model training.
training/pacman_interleaved_cot.jsonl.gz: full-trajectory VQA-style… See the full description on the dataset page: https://huggingface.co/datasets/novastar112/pacman_2d_ultrahard_remap_imagined_rollout_5000.pacman_hard_cot_chunk_k5_train
pacman_hard_cot_chunk_k5_train
BAGEL VLM-Gym world-model dataset (pacman / cot).
CoT chunk-K train set: all-step interleaved imagined reasoning; re-grounds on the true frame 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 pacman checkpoint(s) under the companion model org; CoT and non-CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_hard_cot_chunk_k5_train.pacman_eval_easy_braidthin_random100_v1_20260820
Pacman Easy braided/thinned — random 100 v1
This is the exact 100-episode Pacman Easy shard used by the published UWM trajectories.
Source dataset: pacman_2d_easy_braidthin_oracle_success_v1_20260815 (500 episodes)
Selection: MT19937 random sample without replacement (seed 2026082001)
Records: 100
JSONL SHA-256: 788d71aeb295e3a54ddfca870d8c7c720f59185c43ff21a3105cdc6ecf8d52b0
Full model trajectories:… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_eval_easy_braidthin_random100_v1_20260820.pacman_easy_oracle_success
pacman_easy_oracle_success
BAGEL VLM-Gym world-model dataset (pacman / raw).
Easy Pacman-2D oracle-success rollouts (raw); train/test source for the easy lineage.
layout: Oracle-success rollouts. Gzipped-JSONL shards under data/train/ and data/test/; each row is a full episode (per-step base64-JPEG frames + action strings).
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching pacman checkpoint(s) under the companion model org; CoT and… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_easy_oracle_success.pacman_easy_cot_chunk_k1_train
pacman_easy_cot_chunk_k1_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman checkpoint(s) under the companion model org; CoT and non-CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_easy_cot_chunk_k1_train.pacman_easy_noncot_chunk_k5_train
pacman_easy_noncot_chunk_k5_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman 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/pacman_easy_noncot_chunk_k5_train.pacman_easy_cot_chunk_kinf_train
pacman_easy_cot_chunk_kinf_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman checkpoint(s) under… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_easy_cot_chunk_kinf_train.pacman_braidthin_noncot_chunk_k10_train
pacman_braidthin_noncot_chunk_k10_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman checkpoint(s) under the companion model org; CoT and non-CoT
variants share the… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_braidthin_noncot_chunk_k10_train.pacman_braidthin_nearer_food_eval_500
pacman_braidthin_nearer_food_eval_500
BAGEL VLM-Gym world-model dataset (pacman / eval).
The 500-example easy braided/thinned held-out split used as the nearer-food forward-thinking evaluation.
layout: Test-only oracle-success episodes under data/test/; each row stores the initial state, actions, and base64-JPEG frames.
images are base64-encoded JPEG frames stored inline in each JSONL row.
Pairs with the matching pacman checkpoint(s) under the companion model org; CoT and… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_braidthin_nearer_food_eval_500.pacman_easy_noncot_chunk_k1_train
pacman_easy_noncot_chunk_k1_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman 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/pacman_easy_noncot_chunk_k1_train.pacman_easy_noncot_chunk_k3_train
pacman_easy_noncot_chunk_k3_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman 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/pacman_easy_noncot_chunk_k3_train.pacman_easy_noncot_chunk_k10_train
pacman_easy_noncot_chunk_k10_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman 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/pacman_easy_noncot_chunk_k10_train.pacman_hard_noncot_chunk_k3_train
pacman_hard_noncot_chunk_k3_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman 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/pacman_hard_noncot_chunk_k3_train.pacman_hard_noncot_chunk_k10_train
pacman_hard_noncot_chunk_k10_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman 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/pacman_hard_noncot_chunk_k10_train.pacman_braidthin_noncot_chunk_k1_train
pacman_braidthin_noncot_chunk_k1_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman 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/pacman_braidthin_noncot_chunk_k1_train.pacman_easy_cot_chunk_k3_train
pacman_easy_cot_chunk_k3_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman checkpoint(s) under the companion model org; CoT and non-CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_easy_cot_chunk_k3_train.pacman_easy_cot_chunk_k10_train
pacman_easy_cot_chunk_k10_train
BAGEL VLM-Gym world-model dataset (pacman / cot).
CoT chunk-K train set: all-step interleaved imagined reasoning; re-grounds on the true frame 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 pacman checkpoint(s) under the companion model org; CoT and non-CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_easy_cot_chunk_k10_train.pacman_hard_noncot_chunk_k1_train
pacman_hard_noncot_chunk_k1_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman 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/pacman_hard_noncot_chunk_k1_train.pacman_braidthin_cot_chunk_k3_train
pacman_braidthin_cot_chunk_k3_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman checkpoint(s) under the companion model org; CoT and… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_braidthin_cot_chunk_k3_train.pacman_braidthin_cot_chunk_k5_train
pacman_braidthin_cot_chunk_k5_train
BAGEL VLM-Gym world-model dataset (pacman / cot).
CoT chunk-K train set: all-step interleaved imagined reasoning; re-grounds on the true frame 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 pacman checkpoint(s) under the companion model org; CoT and… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_braidthin_cot_chunk_k5_train.pacman_braidthin_always_forward_kinf_train
pacman_braidthin_always_forward_kinf_train
BAGEL VLM-Gym world-model dataset (pacman / always_forward).
Shared-prefix always-forward K=inf data with a shortest expert continuation and a verified safe longer alternative.
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 pacman checkpoint(s) under the companion… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_braidthin_always_forward_kinf_train.pacman_eval_two_ghost_braidthin_random100_20260817
HF Two-Ghost braided/thinned — deterministic random 100
This is the exact 100-episode Pacman evaluation shard used for the final UWM trajectories.
Source dataset: pacman_2d_easy_g9f8_two_ghost_braidthin_oracle_success_test_v1_20260817
Sampling: Python MT19937 without replacement, seed 2026081702
Records: 100
JSONL SHA-256: fbaf50a9a9799384b2cc1daf5c4c5c7e81c3fed8a23b01cc3b5b53747f3ed034
Trajectories: https://huggingface.co/datasets/novastar111/uwm_paper_eval_trajectories… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_eval_two_ghost_braidthin_random100_20260817.pacman_eval_easy_braidthin_stratified100_v2_20260820
Pacman Easy braided/thinned — stratified 100 v2
This is the exact 100-episode Pacman Easy shard used by the published UWM trajectories.
Source dataset: pacman_2d_easy_braidthin_oracle_success_v1_20260815 (500 episodes)
Selection: oracle-step stratification with distribution balancing (seed 2026082002)
Records: 100
JSONL SHA-256: eadbf33add24bb2e550456a4ad7141e5c111bb32e779df646087533adedd7220
Full model trajectories:… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/pacman_eval_easy_braidthin_stratified100_v2_20260820.pacman_braidthin_noncot_chunk_k3_train
pacman_braidthin_noncot_chunk_k3_train
BAGEL VLM-Gym world-model dataset (pacman / 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 pacman 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/pacman_braidthin_noncot_chunk_k3_train.