Hugging Face's logo

Datasets:
sauravvvv
/
jepa-arcade

Tasks:
Other
Size:
100K<n<1M
Tags:
atari
world-model
jepa
reinforcement-learning
pettingzoo
offline-rl
License:
Dataset card Data Studio Files
xet
 Community
Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
episode_idx: int32
step_idx: int32
action: int32
action_human: int32
pixels: binary
first_x: float
first_y: float
second_x: float
second_y: float
to
{'episode_idx': Value('int32'), 'step_idx': Value('int32'), 'action': Value('int32'), 'action_human': Value('int32'), 'pixels': Image(mode=None, decode=True), 'player_x': Value('float32'), 'player_y': Value('float32'), 'enemy_x': Value('float32'), 'enemy_y': Value('float32'), 'ball_x': Value('float32'), 'ball_y': Value('float32'), 'ball_vx': Value('float32'), 'ball_vy': Value('float32')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1816, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/lance/lance.py", line 231, in _generate_tables
                  yield Key(frag_idx, batch_idx), self._cast_table(table)
                                                  ~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/lance/lance.py", line 188, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              episode_idx: int32
              step_idx: int32
              action: int32
              action_human: int32
              pixels: binary
              first_x: float
              first_y: float
              second_x: float
              second_y: float
              to
              {'episode_idx': Value('int32'), 'step_idx': Value('int32'), 'action': Value('int32'), 'action_human': Value('int32'), 'pixels': Image(mode=None, decode=True), 'player_x': Value('float32'), 'player_y': Value('float32'), 'enemy_x': Value('float32'), 'enemy_y': Value('float32'), 'ball_x': Value('float32'), 'ball_y': Value('float32'), 'ball_vx': Value('float32'), 'ball_vy': Value('float32')}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1869, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

episode_idx
int32
step_idx
int32
action
int32
action_human
int32
pixels
image
player_x
float32
player_y
float32
enemy_x
float32
enemy_y
float32
ball_x
float32
ball_y
float32
ball_vx
float32
ball_vy
float32
0
0
2
9
72
142
72
7
77
2
0
0
0
1
1
2
68
142
72
7
77
11
0
9
0
2
1
3
68
142
72
7
77
18
0
7
0
3
12
9
72
142
72
7
77
22
0
4
0
4
2
3
68
142
68
7
77
25
0
3
0
5
3
10
72
142
68
7
77
26
0
1
0
6
17
1
72
145
72
7
75
21
-2
-5
0
7
16
16
64
145
68
3
73
28
-2
7
0
8
3
1
68
141
64
2
71
34
-2
6
0
9
1
1
68
141
68
2
69
37
-2
3
0
10
1
15
68
141
68
2
67
38
-2
1
0
11
13
1
64
145
68
2
65
37
-2
-1
0
12
6
1
64
145
68
2
63
35
-2
-2
0
13
4
1
64
145
72
6
62
30
-1
-5
0
14
4
1
64
145
68
6
60
23
-2
-7
0
15
15
1
64
145
64
6
58
14
-2
-9
0
16
1
4
64
145
60
10
56
4
-2
-10
0
17
1
1
60
145
60
10
54
5
-2
1
0
18
4
17
60
145
60
10
52
14
-2
9
0
19
1
1
56
141
56
10
50
20
-2
6
0
20
4
4
56
141
56
10
48
24
-2
4
0
21
8
1
52
141
52
10
47
26
-1
2
0
22
4
4
52
141
56
6
45
27
-2
1
0
23
4
4
49
141
53
6
2
25
-43
-2
0
24
4
6
49
141
53
6
2
25
0
0
0
25
14
4
49
141
53
6
2
25
0
0
0
26
4
1
49
141
53
6
2
25
0
0
0
27
6
4
49
141
53
6
2
25
0
0
0
28
4
4
49
141
53
6
2
25
0
0
0
29
4
6
49
141
53
6
2
25
0
0
0
30
17
17
49
141
53
6
2
25
0
0
0
31
0
4
49
141
53
6
2
25
0
0
0
32
4
0
49
141
53
6
2
25
0
0
0
33
0
4
49
141
53
6
2
25
0
0
0
34
17
4
49
141
53
6
2
25
0
0
0
35
9
4
49
141
53
6
2
25
0
0
0
36
4
4
49
141
53
6
2
25
0
0
0
37
4
4
49
141
53
6
2
25
0
0
0
38
16
14
49
141
53
6
2
25
0
0
0
39
9
4
49
141
53
6
2
25
0
0
0
40
4
16
49
141
53
6
2
25
0
0
0
41
11
4
49
141
53
6
2
25
0
0
0
42
0
4
49
141
53
6
2
25
0
0
0
43
0
4
49
141
53
6
2
25
0
0
0
44
11
4
49
141
53
6
2
25
0
0
0
45
16
4
49
141
53
6
2
25
0
0
0
46
4
7
49
141
53
6
2
25
0
0
0
47
10
4
49
141
53
6
2
25
0
0
0
48
12
15
49
141
53
6
2
25
0
0
0
49
4
16
49
141
53
6
2
25
0
0
0
50
4
13
49
141
53
6
2
25
0
0
0
51
4
4
49
141
53
6
2
25
0
0
0
52
4
4
49
141
53
6
2
25
0
0
0
53
4
3
49
141
53
6
2
25
0
0
0
54
4
4
49
141
53
6
2
25
0
0
0
55
3
3
68
142
68
7
77
10
0
-15
0
56
16
1
72
142
72
7
77
17
0
7
0
57
1
1
72
142
76
4
75
21
-2
4
0
58
1
1
72
142
76
4
73
28
-2
7
0
59
1
17
72
142
76
4
72
34
-1
6
0
60
1
1
68
138
76
4
70
37
-2
3
0
61
4
1
68
138
76
4
68
38
-2
1
0
62
1
4
68
138
72
4
66
37
-2
-1
0
63
0
14
64
138
72
4
64
35
-2
-2
0
64
4
12
68
142
72
4
62
30
-2
-5
0
65
6
11
64
142
68
4
60
23
-2
-7
0
66
12
4
68
142
72
8
58
14
-2
-9
0
67
17
4
64
142
68
8
57
4
-1
-10
0
68
4
1
60
142
64
4
55
4
-2
0
0
69
4
4
60
142
60
4
53
13
-2
9
0
70
1
1
56
142
56
4
51
19
-2
6
0
71
4
4
56
142
56
4
49
23
-2
4
0
72
4
1
52
142
52
4
47
25
-2
2
0
73
1
4
52
142
48
4
45
26
-2
1
0
74
4
13
48
142
48
4
2
24
-43
-2
0
75
13
9
48
142
48
4
2
24
0
0
0
76
4
4
48
142
48
4
2
24
0
0
0
77
4
7
48
142
48
4
2
24
0
0
0
78
4
4
48
142
48
4
2
24
0
0
0
79
4
11
48
142
48
4
2
24
0
0
0
80
4
4
48
142
48
4
2
24
0
0
0
81
4
4
48
142
48
4
2
24
0
0
0
82
4
4
48
142
48
4
2
24
0
0
0
83
5
4
48
142
48
4
2
24
0
0
0
84
4
4
48
142
48
4
2
24
0
0
0
85
4
4
48
142
48
4
2
24
0
0
0
86
1
15
48
142
48
4
2
24
0
0
0
87
2
16
48
142
48
4
2
24
0
0
0
88
4
4
48
142
48
4
2
24
0
0
0
89
4
4
48
142
48
4
2
24
0
0
0
90
14
0
48
142
48
4
2
24
0
0
0
91
16
2
48
142
48
4
2
24
0
0
0
92
0
4
48
142
48
4
2
24
0
0
0
93
14
4
48
142
48
4
2
24
0
0
0
94
4
4
48
142
48
4
2
24
0
0
0
95
1
4
48
142
48
4
2
24
0
0
0
96
17
13
48
142
48
4
2
24
0
0
0
97
4
4
48
142
48
4
2
24
0
0
0
98
4
4
48
142
48
4
2
24
0
0
0
99
0
4
48
142
48
4
2
24
0
0
End of preview.

JEPA Arcade — Atari transition datasets

Pixel + RAM-state transition data from two-player PettingZoo Atari environments, collected to train the JEPA Arcade world model.

Code: https://github.com/saurav-34/lepong · Model: sauravvvv/jepa-arcade

Every table is a Lance dataset with PNG-encoded frames and per-frame ground-truth state read from ALE RAM.

Tables

File Env Rows Size State columns
pong_ma_128x128.lance pong_v3 100,000 86 MB player_y, enemy_y, ball_x, ball_y, ball_vx, ball_vy
tennis_ma_128x128_100k.lance tennis_v3 100,000 95 MB player_x/y, enemy_x/y, ball_x/y, ball_vx/vy
boxing_ma_128x128.lance boxing_v2 100,000 259 MB first_x/y, second_x/y
boxing_sample.lance boxing_v2 2,000 5.2 MB same as above

Shared columns: episode_idx, step_idx, action (the JEPA seat), action_human (the reference seat), pixels (PNG-encoded 128×128 RGB).

Collection

All tables come from scripts/collect_pettingzoo.py, driven by one YAML per game. Frames are 128×128 RGB, PNG-encoded; frameskip is 4 throughout.

Policies are a ball-chasing heuristic with ε-random mixed in for action coverage (ε = 0.30 for Pong and Boxing, 0.35 for Tennis).

Two conventions matter if you reuse this data:

  • Frame and RAM state are captured before acting, so (s_t, a_t) is aligned rather than off-by-one.
  • Velocity columns are capped. Real ball motion is ~4–5 RAM px per frameskip-4 step, but serve teleports exceed 70. An uncapped finite difference would record a physics event as a velocity spike, so the cap (24 for Pong, 50 for Tennis) turns teleports into v = 0.

Each chosen action is held for frameskip env steps. If you train on this data, hold the same stride at play time — a mismatch produces a model that correctly predicts the wrong future.

Loading 

import lance
ds = lance.dataset("pong_ma_128x128.lance")
tbl = ds.to_table(limit=8).to_pydict()

import io
from PIL import Image
img = Image.open(io.BytesIO(tbl["pixels"][0]))   # 128×128 RGB

A lance.Dataset handle must never cross fork() — Lance's Tokio runtime does not survive it. Use the spawn start method in DataLoader workers and reopen the dataset inside each worker.

Known issues

  • Tennis ball_y is arc height, not court position. It derives from RAM byte 17, which references label ball_y, but a wiggle-probe shows it tracks the ball's height above the court. Recorded faithfully; treat as suspect.
  • Parked-ball frames are included. ALE parks the ball between points — on Tennis this is ~59% of frames. Train a state probe on them and it earns a high correlation by learning dead-vs-live classification while getting worse at rallies. Filter on ball_x before fitting.
  • Only two seats are stored as action columns. The collector conditions on one seat (action) and stores one reference seat (action_human). For envs with more than two agents the remaining players are still driven, but their actions are not columns.

License

Collection code and annotations are MIT. The frames are rendered from Atari 2600 ROMs via ALE and are derived from copyrighted games; they are released here for non-commercial research use, in line with existing Atari research datasets.
Downloads last month
24
Total file size:
465 MB