Dataset Viewer
Auto-converted to Parquet Duplicate
schedule
stringclasses
1 value
track
stringclasses
1 value
config
stringclasses
1 value
mode
stringclasses
1 value
seed
int64
0
0
users
int64
1
1
step
int64
0
299
task
stringlengths
8
8
phase
stringclasses
2 values
exposure
int64
0
7
gap
int64
1
141
⌀
reset
bool
1 class
input
listlengths
1
30
output
listlengths
1
30
demos
listlengths
2
8
demoCap
int64
2
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
0
88a10436
stream
0
null
false
[ [ 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4 ], [ 4, 4, 1, 4, 4, 4, 4, 4, 4, 1, 4, 4, 4 ], [ 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4 ], [ 4, 4, ...
[ [ 4, 4, 5, 5, 4, 4, 4, 4, 4, 5, 5, 4, 4 ], [ 4, 8, 5, 8, 4, 4, 4, 4, 8, 5, 8, 4, 4 ], [ 4, 8, 5, 0, 4, 4, 4, 4, 8, 5, 0, 4, 4 ], [ 4, 4, ...
[ { "input": [ [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
1
e5062a87
stream
0
null
false
[ [ 5, 0, 5, 0, 5, 0, 0, 0, 0, 5, 5, 0, 0, 0, 0, 5, 0 ], [ 0, 5, 5, 5, 5, 0, 0, 5, 5, 0, 0, 0, 0, 5, 5, 0, 0 ], [ 5, 5, 5, 0, 0, 5, 5, 5, ...
[ [ 5, 0, 5, 0, 5, 0, 0, 0, 0, 5, 5, 1, 0, 0, 0, 5, 1 ], [ 0, 5, 5, 5, 5, 0, 0, 5, 5, 1, 0, 1, 0, 5, 5, 0, 1 ], [ 5, 5, 5, 0, 0, 5, 5, 5, ...
[ { "input": [ [ 5, 5, 5, 0, 0, 0, 5, 5, 5, 0, 5, 5, 5, 5, 5, 5, 0, 0, 5, 5, 5, 5, 5, 5, 0, 5, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
2
e5062a87
stream
1
1
false
[ [ 0, 9, 1, 1, 1, 9, 9, 9, 0, 9, 9, 9, 9, 0, 9, 9, 0, 0, 9, 9, 9, 9, 9, 9, 9 ], [ 9, 9, 9, 0, 1, 9, 9, 9, 0, 9, 0, 9, 9, 9, 9, 9, 9, 9, 9...
[ [ 0, 9, 1, 1, 1, 9, 9, 9, 0, 9, 9, 9, 9, 0, 9, 9, 0, 0, 9, 9, 9, 9, 9, 9, 9 ], [ 9, 9, 9, 0, 1, 9, 9, 9, 0, 9, 0, 9, 9, 9, 9, 9, 9, 9, 9...
[ { "input": [ [ 5, 5, 5, 5, 0, 0, 0, 5, 5, 0, 0, 5, 0 ], [ 5, 5, 5, 0, 0, 5, 0, 0, 0, 5, 5, 5, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
3
e73095fd
stream
0
null
false
[ [ 1, 1, 1, 5, 1, 1, 1, 1, 1, 1, 1, 1, 1, 5, 1, 1, 1, 1, 1, 1, 1, 1, 5, 1 ], [ 1, 1, 1, 5, 1, 1, 1, 1, 1, 1, 1, 1, 1, 5, 1, 1, 1, 1, 1, 1...
[ [ 1, 1, 1, 5, 1, 1, 1, 1, 1, 1, 1, 1, 1, 5, 1, 1, 1, 1, 1, 1, 1, 1, 5, 1 ], [ 1, 1, 1, 5, 1, 1, 1, 1, 1, 1, 1, 1, 1, 5, 1, 1, 1, 1, 1, 1...
[ { "input": [ [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0 ], [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3, 0, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
4
63613498
stream
0
null
false
[ [ 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7 ], [ 7, 7, 7, 7, 7, 7, 7, 7, 7, 0, 7, 7, 7, 3, 3, 7, 7 ], [ 7, 7, 7, 7, 7, 7, 7, 0, ...
[ [ 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7 ], [ 7, 7, 7, 7, 7, 7, 7, 7, 7, 2, 7, 7, 7, 3, 3, 7, 7 ], [ 7, 7, 7, 7, 7, 7, 7, 2, ...
[ { "input": [ [ 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3 ], [ 3, 3, 3, 3, 3, 3, 3, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
5
e73095fd
stream
1
2
false
[ [ 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 8, 2, 2 ], [ 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2...
[ [ 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 8, 2, 2 ], [ 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2...
[ { "input": [ [ 5, 5, 5, 5, 5, 1, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5 ], [ 5, 5, 1, 1, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
6
e73095fd
stream
2
1
false
[ [ 5, 5, 5, 7, 5, 5, 5, 5, 5 ], [ 5, 5, 5, 7, 5, 5, 5, 5, 5 ], [ 7, 7, 7, 7, 7, 5, 5, 5, 5 ], [ 7, 5, 5, 5, 7, 5, 5, 5, 5 ], [ 7, 5, 5, 5, ...
[ [ 5, 5, 5, 7, 5, 5, 5, 5, 5 ], [ 5, 5, 5, 7, 5, 5, 5, 5, 5 ], [ 7, 7, 7, 7, 7, 5, 5, 5, 5 ], [ 7, 4, 4, 4, 7, 5, 5, 5, 5 ], [ 7, 4, 4, 4, ...
[ { "input": [ [ 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7 ], [ 7, 7, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
7
e5062a87
stream
2
5
false
[ [ 0, 9, 0, 0, 9, 0, 0, 0, 9, 9, 9, 9, 0 ], [ 0, 9, 0, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9 ], [ 9, 0, 9, 0, 0, 9, 9, 9, 9, 0, 9, 9, 0 ], [ 9, 9, ...
[ [ 0, 9, 0, 0, 9, 0, 0, 0, 9, 9, 9, 9, 0 ], [ 0, 9, 0, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9 ], [ 9, 0, 9, 5, 0, 9, 9, 9, 9, 0, 9, 9, 0 ], [ 9, 9, ...
[ { "input": [ [ 0, 3, 3, 3, 3, 3, 0, 3, 3, 3, 3, 3, 3, 3, 3, 0, 0, 3, 3, 3, 0, 3, 3, 3 ], [ 3, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
8
e5062a87
stream
3
1
false
[ [ 6, 6, 0, 6, 0, 6, 6, 6, 6, 0, 6 ], [ 6, 6, 6, 6, 0, 0, 6, 0, 0, 0, 6 ], [ 6, 6, 6, 6, 0, 6, 8, 8, 6, 6, 6 ], [ 0, 0, 6, 6, 0, 6, 8, 0, ...
[ [ 6, 6, 0, 6, 0, 6, 6, 6, 6, 0, 6 ], [ 6, 6, 6, 6, 8, 8, 6, 0, 0, 0, 6 ], [ 6, 6, 6, 6, 8, 6, 8, 8, 6, 6, 6 ], [ 8, 8, 6, 6, 0, 6, 8, 0, ...
[ { "input": [ [ 9, 9, 0, 9, 9, 9, 9, 9, 9, 0, 9, 9, 9, 9, 9, 9, 9, 0, 9, 9 ], [ 9, 9, 9, 9, 9, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
9
e73095fd
stream
3
3
false
[ [ 3, 3, 3, 3, 3, 6, 3, 3, 3, 3, 3, 3, 3, 3 ], [ 3, 3, 3, 3, 3, 6, 3, 3, 3, 3, 3, 3, 3, 3 ], [ 3, 3, 3, 3, 3, 6, 3, 3, 3, 3, 3, 3, 3, 3 ], ...
[ [ 3, 3, 3, 3, 3, 6, 3, 3, 3, 3, 3, 3, 3, 3 ], [ 3, 3, 3, 3, 3, 6, 3, 3, 3, 3, 3, 3, 3, 3 ], [ 3, 3, 3, 3, 3, 6, 3, 3, 3, 3, 3, 3, 3, 3 ], ...
[ { "input": [ [ 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8 ], [ 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
10
6773b310
stream
0
null
false
[ [ 6, 6, 6, 5, 3, 6, 6, 6, 5, 3, 6, 6, 6, 5 ], [ 5, 6, 6, 6, 3, 6, 6, 6, 6, 3, 6, 6, 5, 6 ], [ 6, 6, 6, 6, 3, 6, 6, 6, 6, 3, 6, 6, 6, 6 ], ...
[ [ 1, 6, 1 ], [ 6, 1, 6 ], [ 1, 1, 1 ] ]
[ { "input": [ [ 2, 2, 7, 2, 9, 7, 2, 9, 7, 9, 2, 7, 2, 2 ], [ 9, 2, 7, 2, 2, 7, 2, 2, 7, 2, 9, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
11
673ef223
stream
0
null
false
[ [ 5, 6, 6, 6, 6, 6 ], [ 5, 6, 6, 6, 6, 6 ], [ 5, 6, 6, 6, 6, 6 ], [ 5, 6, 6, 6, 6, 6 ], [ 6, 6, 6, 6, 6, 6 ], [ 6, 6, 6, 6, 6, 6 ], [ 6, 2, ...
[ [ 5, 2, 2, 2, 2, 2 ], [ 5, 2, 2, 2, 2, 2 ], [ 5, 2, 2, 2, 2, 2 ], [ 5, 2, 2, 2, 2, 2 ], [ 6, 6, 6, 6, 6, 6 ], [ 6, 6, 6, 6, 6, 6 ], [ 6, 4, ...
[ { "input": [ [ 1, 3, 3, 1, 1, 1, 1 ], [ 1, 1, 1, 1, 1, 1, 1 ], [ 1, 1, 1, 1, 1, 1, 1 ], [ 1, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
12
f76d97a5
stream
0
null
false
[ [ 9, 9, 9, 9, 9, 5, 9, 9, 5, 9, 9, 5, 9, 5, 9, 5, 5, 9, 5, 9, 9, 5, 5, 9, 5, 9, 9, 9, 5, 9 ], [ 5, 9, 5, 9, 9, 9, 9, 9, 9, 9, 5, 9, 5, 9...
[ [ 0, 0, 0, 0, 0, 9, 0, 0, 9, 0, 0, 9, 0, 9, 0, 9, 9, 0, 9, 0, 0, 9, 9, 0, 9, 0, 0, 0, 9, 0 ], [ 9, 0, 9, 0, 0, 0, 0, 0, 0, 0, 9, 0, 9, 0...
[ { "input": [ [ 5, 7, 7, 5, 7, 5, 7, 5, 7, 5, 5, 5, 5, 7, 5, 5, 7, 5, 5, 5, 5 ], [ 5, 5, 5, 5, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
13
63613498
stream
1
9
false
[ [ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 5, 8, 8, 8, 5, 5 ], [ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5...
[ [ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 5, 8, 8, 8, 5, 5 ], [ 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5...
[ { "input": [ [ 3, 3, 3, 3, 3, 4, 4, 3, 5, 5, 3, 3, 3, 3, 2, 3, 5, 5, 3 ], [ 3, 3, 3, 3, 3, 4, ...
8
arc1-small-workstreams0
arc1
arc1-small
workstreams
0
1
14
673ef223
stream
1
3
false
[ [ 5, 5, 5, 5, 5, 5, 0 ], [ 5, 5, 5, 5, 5, 5, 0 ], [ 5, 5, 5, 5, 5, 5, 0 ], [ 5, 5, 5, 5, 5, 5, 5 ], [ 0, 5, 5, 7, 5, 5, 5 ], [ 0, 5, 5, 7, ...
[ [ 7, 7, 7, 7, 7, 7, 0 ], [ 7, 7, 7, 7, 7, 7, 0 ], [ 5, 5, 5, 5, 5, 5, 0 ], [ 5, 5, 5, 5, 5, 5, 5 ], [ 0, 7, 7, 4, 5, 5, 5 ], [ 0, 7, 7, 4, ...
[ { "input": [ [ 2, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8 ], [ 2, 8, 8, ...
8
End of preview. Expand in Data Studio

Continual-ARC

The instances and tasks of Continual-ARC: Measuring Skill Acquisition over a Working Lifetime (Daniel Lenton, Yusha Arif and Haris Mahmood, 2026): the instances the paper's runs were served on ARC-AGI-1 and ARC-AGI-2 and the levels they were served on ARC-AGI-3, each schedule's in the order it served them, and the 125 tasks of the ARC-AGI-1 track with their own public pairs. The benchmark's code, the runs' records and trajectories, and these files as plain JSON are in unifyai/continual-arc (data/), and continual-arc.unify.ai shows them. The paper's page here is huggingface.co/papers/2611.14505.

In Continual-ARC the same tasks return after dormant gaps with fresh instances, while new tasks keep arriving, and the learner pays for each demonstration it buys and each wrong answer it submits. After the stream, the learner is frozen and tested on fresh instances of the tasks it met and on unseen tasks.

Loading

from datasets import load_dataset

# one schedule's instances, in the order it served them
stream = load_dataset("unifyai/continual-arc", split="arc1_small_workstreams0")

# the ARC-AGI-3 stream's levels
levels = load_dataset("unifyai/continual-arc", "arc3", split="arc3_workstreams0")

# the 125 tasks of the ARC-AGI-1 track
tasks = load_dataset("unifyai/continual-arc", "tasks", split="arc1")

instances

A row per instance and a split per schedule, named as the schedule is with underscores for its hyphens, its rows in the order the schedule served them.

Split Instances Served to
arc1_small_workstreams0 300 the runs in the track's own order at seed 0 (ws0 in their names): its stream, then the epilogue, the held-out tasks once each while the learner still learns, which no measure reads
arc1_small_sequential0, arc1_small_random0 200 each the runs in the two other orders of the paper's Appendix B (seq, rnd)
arc1_small_fresh0 100 the held-out tasks alone, the learner wiped before each (fresh)
arc1_small_workstreams0_u2, arc1_small_workstreams0_u4 400, 800 the open-model runs of two and four users
arc2_workstreams0 200 the runs on ARC-AGI-2 (the paper's Appendix K)
arc2_small_workstreams0 200 the same stream, slot for slot and instance for instance, under the name the demonstration-cache runs KC and KC0 were played under
Column What it holds
schedule, track, config, mode, seed, users the schedule that served the instance: its name, its track (arc1 or arc2), configuration, order, seed and how many users it served
step the instance's place in the schedule, from 0
task, phase its task, and its phase, stream or epilogue
exposure, gap, reset how many times the schedule had served the task before, the steps since it last did (empty the first time), and whether the learner was reset before the instance
input, output the test input and its right output, as grids of colors 0 to 9
demos every demonstration the instance offers, input-output pairs in the order a learner buys them, at most demoCap

The runs name the first track's configuration arc1-small, as every run played before v0.6.0 does, when the benchmark still named its configurations by size; its streams are the ones arc1 builds today.

arc3

ARC-AGI-3's levels, a row per level and a split per schedule, its rows in the order the schedule served them. An instance of this track is a fresh level of a game's first level, drawn by the benchmark's own generator for the game and solved by its solver. The learner plays it in ARC Prize's game, sees frames of 64x64 cells in colors 0 to 15, and can buy up to three demonstrations, the solver playing other fresh levels of the same game.

Split Levels Served to
arc3_workstreams0 120 the runs on ARC-AGI-3 (the paper's Appendix K): eight levels of each of 15 games, in the track's own order at seed 0; the demonstration-cache runs KC and KC0 played the same stream under the name arc3-small-workstreams
arc3_seen_test0 60 the seen test the runs played after the stream, every level from the learner's state at the stream's end: four new levels of each of the stream's 15 games
arc3_held_out_test0 40 the held-out test, played the same way: four levels of each of the 10 games the stream never serves
arc3_small_fresh0 40 the held-out test's levels alone, the model wiped before each: the first-sight runs
Column What it holds
schedule, track, config, mode, seed, users the schedule that served the level, as in instances
step the level's place in the schedule, from 0
game, version the game, by ARC Prize's id, and the version of it the benchmark pins
phase, exposure, gap, reset as in instances; the phase is stream, seen-test, held-out-test, or epilogue for the held-out levels in the fresh order
instanceSeed the seed the benchmark draws the level from
frame the level's first frame: 64 rows of 64 colors, top row first
availableActions the actions the game accepts: 1 to 5 and 7 simple, 6 a click
baseline the actions a person needs for the level: ARC Prize's count for the game's level 0, scaled by the solver's solution length for this level over level 0's
actionCap five times the baseline, rounded up: every action and reset counts toward it, and a level that reaches it unsolved costs the whole budget
solution the solver's winning actions, each [id], or [6, x, y] for a click at column x and row y
demos the three demonstrations a learner can buy, in the order bought, each the solver playing another fresh level of the game to its end: its first frame (start) and each action with the cells it changed (steps, a change [x, y, color])
demoCap the most demonstrations a level offers, 3

To play a level, install the benchmark with its arc3 extra, fetch ARC Prize's games at their pinned versions (uv run continual-arc data fetch --datasets arc3), and draw the level from its game and seed (LevelSource().instance(game, instanceSeed), in continual_arc.arc3.instances), as the benchmark's environment does.

tasks

The 125 tasks of the ARC-AGI-1 track the GPT-6 Luna runs played, a row each: the task, the ARC-AGI-1 split it comes from, the phases that serve it, stream for the 25 tasks of the stream, from ARC-AGI-1's training set, and epilogue for the 100 ARC-AGI-1 evaluation tasks of the held-out test, and its own public pairs as ARC-AGI-1 publishes them (train, test).

How the instances were made

The instances were written by serving each schedule through the benchmark's own environment (src/continual_arc/env.py) with an oracle that buys every demonstration an instance offers and then submits the right answer. The environment draws each instance and each demonstration from its slot's seed alone, so these are exactly the pairs every learner on that schedule saw, and the benchmark draws them again from the same schedule with the upstream data fetched (uv run continual-arc data fetch). Other seeds draw other streams from the same process.

ARC-AGI-3's levels were drawn again from each slot's seed by the benchmark's own LevelSource (src/continual_arc/arc3/instances.py), the source its environment serves them from, and checked against every run on the track: the baseline, action cap and solution length in each of their 2,320 records, the first board in each of the 1,426 transcripts that kept it, cell for cell, and the first board and length of each of the 2,102 demonstrations learners bought. Playing a level needs its game: ARC Prize's games are MIT-licensed Python on arcengine, which the benchmark downloads at their pinned versions.

License

The grids come from ARC-AGI-1 and ARC-AGI-2 (Apache-2.0) and from the generators that draw their fresh instances, Re-ARC's (MIT) and ARC-GEN's (Apache-2.0); each keeps its source's license. ARC-AGI-3's levels are drawn by the benchmark's own generators (Apache-2.0), and their frames are the renderings of ARC Prize's games (MIT), which playing a level needs.

Citation

@article{lenton2026continualarc,
  title   = {Continual-ARC: Measuring Skill Acquisition over a Working Lifetime},
  author  = {Lenton, Daniel and Arif, Yusha and Mahmood, Haris},
  journal = {arXiv preprint arXiv:2611.14505},
  year    = {2026}
}
Downloads last month
99

Paper for unifyai/continual-arc