Simulated measurement records of the buffer of a dissipative cat qubit, for the Alice & Bob challenge "Catch the flip" at the PushQuantum Hackathon (Munich, 20–22 November 2026). Tier A holds the click records of a photon counter, and it is the tutorial dataset of the challenge. In these records a flip can be seen by eye: tutorial 5 of the workspace Alice_Bob/catch-the-flip uses them to go through the job of a flip detector once, and they help with the physics questions of the challenge. They are not scored, and nothing about them is submitted. The records of the leaderboard are the heterodyne records, the dataset Alice_Bob/catch-the-flip-tier-b. The statement of the challenge, the tutorials and a template notebook are in the workspace.
Explore the data
The table is the result of the SQL query in the box below it. The dataset is the table data: edit the query to filter, count or sort. The table under the box counts the segments and the flips of each split and cat size (the labels of the test split are hidden).
Showing the first 10 of 75,000 rows. Increase the table's page_size to embed more in the static export.
split
str
segment_id
i32
nbar
f64
label
i8
clicktimes
list[f64]
train
0
3.5
0
train
1
3.5
0
train
2
3.0
1
train
3
2.5
0
train
4
3.5
0
train
5
3.0
0
train
6
3.5
0
train
7
3.5
0
train
8
3.0
0
train
9
3.5
0
75,000 rows, 5 columns
No selection
No selection
SELECT * FROM data
Showing the first 10 of 12 rows. Increase the table's page_size to embed more in the static export.
split
str
nbar
f64
segments
i64
flips
i32
flip_percent
f64
train
2.0
2,000
289
14.45
train
2.5
4,000
290
7.25
train
3.0
8,000
297
3.71
train
3.5
16,000
232
1.45
val
2.0
1,000
243
24.30
val
2.5
2,000
296
14.80
val
3.0
4,000
353
8.82
val
3.5
8,000
365
4.56
test
2.0
2,000
None
None
test
2.5
4,000
None
None
12 rows, 5 columns
No selection
No selection
What a row is
A row is one segment, one memory experiment. The cat is prepared in the state +α at t = 0, the output of its buffer is monitored for a duration T_seg = 200, and the cat is read out in the Z basis. Times are in units of 1/κ2, the inverse of the two-photon dissipation rate. The label is 1 when the readout gave −α: the cat has flipped.The record of a segment is the list of the click times of a photon counter on the buffer output. A detection efficiency of 80 % and dark counts of rate 0.02 per unit time, uniform in the segment, are already applied. The simulation resolves a click to one integration step (0.001 time units); click times are spread uniformly within that step, so real clicks and dark counts look alike. A segment holds 12 to 19 clicks on average, depending on the device and the cat size.Every split mixes segments of four cat sizes, nbar = 2, 2.5, 3 and 3.5 photons, in the proportions 1 : 2 : 4 : 8 (big cats flip less often, so they get more segments), in random order. The column nbar gives the cat size of each segment.
Splits
Each split is one partition of the dataset.
split
partition
device
labels
segments
use
train
0
nominal (the one you can simulate)
yes
30 000
tutorial 5; the burst and what it tells you (steps 2 and 4 of the challenge)
val
1
hidden test device
yes
15 000
tutorial 5; the hidden device seen by a photon counter
test
2
hidden test device
no
30 000
not scored: more unlabelled records of the hidden device, if you want them
The hidden device differs from the nominal one in kind, not only in numbers: its buffer is a real resonator of finite bandwidth (two-mode model) and its memory has a small Kerr non-linearity, the same at every cat size. The noise rates and the cat sizes are the nominal ones.Fraction of flip labels:
cat size nbar
2
2.5
3
3.5
nominal device, master equation
14.6 %
7.30 %
3.40 %
1.51 %
train
14.4 ± 0.8 %
7.2 ± 0.4 %
3.7 ± 0.2 %
1.45 ± 0.09 %
val (hidden device)
24.3 ± 1.4 %
14.8 ± 0.8 %
8.8 ± 0.4 %
4.6 ± 0.2 %
The errors are binomial.
Columns
column
type
description
split
string
train, val or test
segment_id
int32
position of the segment in its split, starting at 0
nbar
float64
cat size of the segment (mean photon number of the coherent states)
label
int8
1 if the cat has flipped, 0 otherwise; null in the test split
clicktimes
list of float64
click times of the photon counter, sorted, between 0 and 200
Loading
importnumpyasnpimportpyarrow.computeaspcfromaqora.pyarrowimportdataset# pip install "aqora[pyarrow]"ds=dataset("Alice_Bob/catch-the-flip-tier-a","v1.0.0")train=ds.to_table(filter=pc.field("split")=="train")nbar=train["nbar"].to_numpy()labels=train["label"].to_numpy()clicks=train["clicktimes"].combine_chunks()counts=clicks.value_lengths().to_numpy()# number of clicks of each segmenttimes=clicks.flatten().to_numpy()# all click times, one segment after the other
The helper package catchtheflip of the challenge works with one array of shape (segments, columns), padded with NaN. In the workspace, catchtheflip.records.load_aqora("A", "train") returns the split in that layout, as a dictionary of NumPy arrays (nbar, labels, clicktimes, meta). By hand:
With pandas: pd.read_parquet("aqora://Alice_Bob/catch-the-flip-tier-a/v1.0.0", filters=[("split", "==", "train")]).
Not scored
There is no submission and no leaderboard for the click records. The challenge is scored on the heterodyne records of the dataset Alice_Bob/catch-the-flip-tier-b, whose page describes the submission format. To measure a detector of your own on the click records, use the labelled val split: catchtheflip.scoring.rho_at_yield(scores, labels) computes the score of the leaderboard, as in tutorial 5.
Make your own data
The script scripts/generate_dataset.py of the workspace produces labelled data of the nominal device at any cat size and with any seed. The test split cannot be regenerated: its seed and the parameters of its device are private.