Alice_Bob / Catch the flip, tier B: heterodyne detection
data: edit the query to filter, count or sort. The query below shows the first 6 of the 400 bins of I and Q; SELECT * FROM data shows the full records. 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).page_size to embed more in the static export.SELECT split, segment_id, nbar, label,
I[1:6] AS I_first_6_bins, Q[1:6] AS Q_first_6_bins
FROM datapage_size to embed more in the static export.nbar gives the cat size of each segment.| split | partition | device | labels | segments | use |
|---|---|---|---|---|---|
train | 0 | nominal (the one you can simulate) | yes | 15 000 | design and train your detector |
val | 1 | hidden test device | yes | 7 500 | measure how your detector transfers to the real device |
test | 2 | hidden test device | no | 22 500 | your submission, scored on the leaderboard |
| cat size nbar | 2 | 2.5 | 3 | 3.5 |
|---|---|---|---|---|
| nominal device, master equation | 14.6 % | 7.30 % | 3.40 % | 1.51 % |
train | 15.8 ± 1.2 % | 6.7 ± 0.6 % | 3.0 ± 0.3 % | 1.59 ± 0.14 % |
val (hidden device) | 26.2 ± 2.0 % | 11.3 ± 1.0 % | 8.4 ± 0.6 % | 4.3 ± 0.3 % |
| column | type | description |
|---|---|---|
split | string | train, val or test |
segment_id | int32 | position of the segment in its split, starting at 0; the segment_id of a submission is the one of the test split |
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 |
I | list of 400 float32 | quadrature I of the buffer output, one value per bin |
Q | list of 400 float32 | quadrature Q of the buffer output, one value per bin |
I and Q have the precision of 16-bit floats (a resolution of about 0.001, far below the noise).import pyarrow.compute as pc
from aqora.pyarrow import dataset # pip install "aqora[pyarrow]"
ds = dataset("Alice_Bob/catch-the-flip-tier-b", "v1.0.0")
train = ds.to_table(filter=pc.field("split") == "train")
nbar = train["nbar"].to_numpy()
labels = train["label"].to_numpy()
I = train["I"].combine_chunks().flatten().to_numpy().reshape(len(train), -1) # shape (segments, 400)
Q = train["Q"].combine_chunks().flatten().to_numpy().reshape(len(train), -1)
pd.read_parquet("aqora://Alice_Bob/catch-the-flip-tier-b/v1.0.0", filters=[("split", "==", "train")]). The three splits take 150 MB in memory as arrays: load one split at a time.
In the workspace of the challenge, catchtheflip.records.load_aqora("B", "train") returns the split as a dictionary of NumPy arrays (nbar, labels, I, Q, meta), with I and Q as 16-bit floats.segment_id:
segment_id,score,decision
0,0.12,0
1,3.41,1
...
score is any real number, where higher means more suspicious. The leaderboard discards the 30 % of the segments with the highest scores, all cat sizes together, and reports the flip suppression, 1 − rho70, where rho70 is the fraction of flips among the segments that are kept divided by the fraction of flips among all segments. decision is 0 or 1 (1 = discard) and marks the segments that you yourself would discard. To get your score, save the file from the template notebook of the workspace, then sign in on the live leaderboard and press Submit.
The leaderboard shows the score with two decimals. A team may submit as often as it likes, and its final submission is one file: the last one it submitted before the deadline.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.