datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
dyck-k128-seq_len_2048-1B
dyck-k128-seq_len_2048-1B
Procedurally generated k-shuffle Dyck bracket sequences (Hu et al. 2025, arXiv:2502.19249), as flat uint16 token-id .bin files. Token ids are 0-based: opening bracket type i is id i and its matching close is i + k, so ids span [0, 2k) and the vocabulary is 2k = 256.
Grammar parameters
param
value
k (bracket types)
128
max_depth
16
p_open
0.5
seq_length
2048
file
split
tokens
train.bin
train
999,999,488
val.bin
val
10,000… See the full description on the dataset page: https://huggingface.co/datasets/alexkstern/dyck-k128-seq_len_2048-1B.rr_three_tasks_v1
rr_three_tasks_v1
Three tasks on a Trossen AI solo arm, merged into one LeRobot v2.1 dataset.
task
episodes
frames
pick_specific_item_from_clutter
243
59088
pick_two_in_order
99
40478
open_pot_and_place
100
47288
meta/sources.jsonl maps every episode to its source dataset, episode and revision, with the
staging record (open_pot_and_place variant, pick_two second object, sheet row).
Held-out evaluation episodes
meta/eval_episodes_v1.json: 44… See the full description on the dataset page: https://huggingface.co/datasets/k1seul/rr_three_tasks_v1.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.dapo-math-17k-difficulty-qwen3-1.7b-base-k16
DAPO-Math-17k difficulty under Qwen3-1.7B-Base (K=16)
For each of the 17,398 problems in the DAPO-Math-17k train set, how many of
K=16 samples from the untrained base model are correct.
The headline: 57.27% of problems are solved 0 out of 16 times, and not one
problem is solved 16 out of 16. Difficulty here is entirely one-sided.
Why count per problem instead of reporting mean accuracy
In group-relative RL (GRPO and its relatives), a prompt group whose K responses… See the full description on the dataset page: https://huggingface.co/datasets/RyanYr/dapo-math-17k-difficulty-qwen3-1.7b-base-k16.sokoban_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.pusht_noncot_chunk_k10_train
pusht_noncot_chunk_k10_train
BAGEL VLM-Gym world-model dataset (pusht / 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 pusht 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/pusht_noncot_chunk_k10_train.pusht_96_norm4_cot_chunk_k10_20260622_perseg
pusht_96_norm4_cot_chunk_k10_20260622_perseg
PushT (96px, norm4, JPEG q90; coverage task, no hard split — in-dist claims only) — action-conditioned visual world-model SFT data (CoT self-rollout) for the
BAGEL-7B-MoT feedback-interval study.
Format: gzipped JSONL shards under training/, 1 row = 1 packed episode. CoT rows: per-segment
layout — <think> per-step imagined frame (MSE target) </think> + committed action chunk, with a
loss-0 "Action executed." + real frame re-grounding… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/pusht_96_norm4_cot_chunk_k10_20260622_perseg.code_rose_initial_1_7B_SFT_10K_rollouts_Qwen3-4B-Thinking-2507_k12_t0.7_maxtok12288
code_rose_initial_1_7B_SFT_10K — rollouts (Qwen3-4B-Thinking-2507, k=12)
Pass@k completions generated with vLLM over the prefixes in
CL-From-Nothing/code_rose_initial_1_7B_SFT_10K.
Generation config
Model
Qwen3-4B-Thinking-2507
Samples per question (k)
12
Temperature
0.7
top_p
0.9
max_tokens
12288
max_model_len
32768
Questions
7250 (index 0–7249, full split)
Total rows
87000 (7250 × 12)
Generated by complete_prefix_vllm.py… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/code_rose_initial_1_7B_SFT_10K_rollouts_Qwen3-4B-Thinking-2507_k12_t0.7_maxtok12288.maze2d_easy_noncot_chunk_k10_train
maze2d_easy_noncot_chunk_k10_train
BAGEL VLM-Gym world-model dataset (maze2d / 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 maze2d 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/maze2d_easy_noncot_chunk_k10_train.maze2d_easy_native256_noncot_chunk_k10_20260707_perseg
maze2d_easy_native256_noncot_chunk_k10_20260707_perseg
Maze2d (native 256px, JPEG q95; navigation with stop-required success, easy→hard split) — action-conditioned visual world-model SFT data (non-CoT action-chunk baseline) for the
BAGEL-7B-MoT feedback-interval study.
Format: gzipped JSONL shards under training/, 1 row = 1 packed episode. CoT rows: per-segment
layout — <think> per-step imagined frame (MSE target) </think> + committed action chunk, with a
loss-0 "Action… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/maze2d_easy_native256_noncot_chunk_k10_20260707_perseg.NMC-cn_k12-20k-r1_32b_distilled本数据集数据来源为NuminaMath-CoT数据集的cn_k12数据。我们从这里面提取了20000条问题，并使用DeepSeek-R1-Distill-Qwen-32B模型进行了回答。
distilled_s0_e20000.jsonl包含这个数据集的数据，下面介绍数据标签：
idx：索引号（0~19999）
question：原数据集中的problem标签，是一个可能包含多个子问题的数学问题字符串
gt_cot：愿数据集中的solution标签，是经过GPT-4o整理的答案字符串
pred_cot：根据question标签，模型DeepSeek-R1-Distill-Qwen-32B的回答字符串
pred_cot_token_len：pred_cot标签下的字符串转化成token之后的长度（不包含最前面的<think>\n部分，这个在生成的时候是在prompt里面，我后来加到这里了）
message：根据question标签和pred_cot标签，构造的问题-回答数据对
统计了一下平均回答token长度，为3169.4251
pusht_96_norm4_noncot_chunk_k10_20260622_perseg
pusht_96_norm4_noncot_chunk_k10_20260622_perseg
PushT (96px, norm4, JPEG q90; coverage task, no hard split — in-dist claims only) — action-conditioned visual world-model SFT data (non-CoT action-chunk baseline) for the
BAGEL-7B-MoT feedback-interval study.
Format: gzipped JSONL shards under training/, 1 row = 1 packed episode. CoT rows: per-segment
layout — <think> per-step imagined frame (MSE target) </think> + committed action chunk, with a
loss-0 "Action executed." + real… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/pusht_96_norm4_noncot_chunk_k10_20260622_perseg.maze2d_easy_native256_cot_chunk_k10_20260707_perseg
maze2d_easy_native256_cot_chunk_k10_20260707_perseg
Maze2d (native 256px, JPEG q95; navigation with stop-required success, easy→hard split) — action-conditioned visual world-model SFT data (CoT self-rollout) for the
BAGEL-7B-MoT feedback-interval study.
Format: gzipped JSONL shards under training/, 1 row = 1 packed episode. CoT rows: per-segment
layout — <think> per-step imagined frame (MSE target) </think> + committed action chunk, with a
loss-0 "Action executed." + real frame… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/maze2d_easy_native256_cot_chunk_k10_20260707_perseg.maze2d_easy_cot_chunk_k10_train
maze2d_easy_cot_chunk_k10_train
BAGEL VLM-Gym world-model dataset (maze2d / 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 maze2d checkpoint(s) under the companion model org; CoT and non-CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/maze2d_easy_cot_chunk_k10_train.pusht_96_norm4_cot_chunk_k1_20260622_perseg
pusht_96_norm4_cot_chunk_k1_20260622_perseg
PushT (96px, norm4, JPEG q90; coverage task, no hard split — in-dist claims only) — action-conditioned visual world-model SFT data (CoT self-rollout) for the
BAGEL-7B-MoT feedback-interval study.
Format: gzipped JSONL shards under training/, 1 row = 1 packed episode. CoT rows: per-segment
layout — <think> per-step imagined frame (MSE target) </think> + committed action chunk, with a
loss-0 "Action executed." + real frame re-grounding… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/pusht_96_norm4_cot_chunk_k1_20260622_perseg.sokoban_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.maze2d_easy_native256_noncot_chunk_k1_20260707_perseg
maze2d_easy_native256_noncot_chunk_k1_20260707_perseg
Maze2d (native 256px, JPEG q95; navigation with stop-required success, easy→hard split) — action-conditioned visual world-model SFT data (non-CoT action-chunk baseline) for the
BAGEL-7B-MoT feedback-interval study.
Format: gzipped JSONL shards under training/, 1 row = 1 packed episode. CoT rows: per-segment
layout — <think> per-step imagined frame (MSE target) </think> + committed action chunk, with a
loss-0 "Action… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/maze2d_easy_native256_noncot_chunk_k1_20260707_perseg.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.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.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.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.maze2d_easy_cot_chunk_k1_train
maze2d_easy_cot_chunk_k1_train
BAGEL VLM-Gym world-model dataset (maze2d / 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 maze2d checkpoint(s) under the companion model org; CoT and non-CoT… See the full description on the dataset page: https://huggingface.co/datasets/novastar111/maze2d_easy_cot_chunk_k1_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_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_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_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.maze2d_easy_native256_cot_chunk_k1_20260707_perseg
maze2d_easy_native256_cot_chunk_k1_20260707_perseg
Maze2d (native 256px, JPEG q95; navigation with stop-required success, easy→hard split) — action-conditioned visual world-model SFT data (CoT self-rollout) for the
BAGEL-7B-MoT feedback-interval study.
Format: gzipped JSONL shards under training/, 1 row = 1 packed episode. CoT rows: per-segment
layout — <think> per-step imagined frame (MSE target) </think> + committed action chunk, with a
loss-0 "Action executed." + real frame… See the full description on the dataset page: https://huggingface.co/datasets/ultrastar111/maze2d_easy_native256_cot_chunk_k1_20260707_perseg.maze2d_easy_noncot_chunk_k1_train
maze2d_easy_noncot_chunk_k1_train
BAGEL VLM-Gym world-model dataset (maze2d / 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 maze2d 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/maze2d_easy_noncot_chunk_k1_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.pusht_noncot_chunk_k1_train
pusht_noncot_chunk_k1_train
BAGEL VLM-Gym world-model dataset (pusht / 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 pusht 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/pusht_noncot_chunk_k1_train.