ShayManor/willow-surface-code-detection-events

Willow Surface-Code Detection Events (ingested)

Detection events and logical-observable flips derived from Google's Willow below-threshold surface-code dataset (Zenodo 10.5281/zenodo.13273331), rotated surface code at distances 3, 5, 7 in X and Z memory.

Each row is one experimental shot. Detection events are derived from the raw device measurement records with Stim's measurement-to-detector converter, using the per-shot sweep bits, and validated to reproduce the dataset's shipped detection_events.b8 / obs_flips_actual.b8 byte-for-byte for all 420 configurations.

Columns

column	type	meaning
distance	int16	code distance d (3, 5, 7)
basis	string	X or Z memory
rounds	int16	number of QEC cycles
orientation	string	patch location label on the chip (e.g. q4_5)
shot	int32	shot index within the configuration
detectors	list<bool>	which detectors fired (length num_detectors, = 2 * rounds * (d^2-1)/2)
observable	bool	ground-truth logical flip the decoder must predict

One Parquet shard per (distance, orientation, basis, rounds) configuration under data/. The decoder sees only detectors; observable is the held-out answer key.

Decoding bundle

Alongside the detection events, each config ships the inputs a decoder panel needs, keyed by the same <stem> (d{D}_at_{orient}__{basis}__r{rounds:03d}):

• circuits/<stem>.stim — the ideal (noiseless) annotated circuit.
• dems/<stem>.si1000.dem.gz — shipped SI1000 detector error model (gzipped).
• dems/<stem>.rl.dem.gz — shipped RL-optimized detector error model (gzipped).

With these, the whole evaluation runs off this dataset with no local copy of the 12 GB Willow tree.

Load

from datasets import load_dataset
ds = load_dataset("ShayManor/willow-surface-code-detection-events", split="train")     # detection events