atal-narayan/copra-flow-tg-frog

COPRA Flow TG-FROG

A synthetic dataset for learning to refine ultrafast pulse reconstructions from transient-grating frequency-resolved optical gating (TG-FROG) traces. It contains 10,000 pulses, each paired with 30 reconstruction candidates: 300,000 source–target pairs in total.

• Code: atalnarayan/copra-flow
• Trained model: atal-narayan/copra-flow
• License: CC BY 4.0
• Size: approximately 3.8 GB to download; approximately another 3.8 GB to extract

Download and use

The browser dataset viewer is disabled because the pulse fields contain complex-valued NumPy arrays that its table conversion does not support. Download and extract the NPZ shards using the instructions below.

Install the Hugging Face CLI, then download and extract the dataset:

hf download atal-narayan/copra-flow-tg-frog --repo-type dataset \
  --revision v0.1.1 --local-dir data/flow_matching_10k
python data/flow_matching_10k/extract.py

The extraction script verifies each archive and every extracted pulse against SHA-256 checksums. It creates data/flow_matching_10k/shards/, ready for the training and evaluation commands in the code repository. Keep splits.json with the dataset. The tar files can be removed after successful extraction if you need the disk space.

Download and extract the complete dataset for the supplied training and benchmark commands: they validate membership against the full fixed split. This is a scientific NumPy dataset, loaded through the project code or numpy.load; it is not a tabular datasets.load_dataset collection.

What is included

Files	Contents
archives/train-000.tar … train-008.tar	9,000 training pulse shards
archives/val-000.tar	1,000 validation pulse shards
splits.json	Exact pulse UID membership for both partitions
manifest.json	Pulse families, parameters, UIDs, and generation seeds
generation_config.json	Settings used to generate the data
inventory.json	Per-shard and per-archive membership, sizes, and checksums
array_schema.json, DATA_FORMAT.md	Array layouts, units, and normalization conventions
SHA256SUMS, SHARD_SHA256SUMS	Checksums for release files and extracted shards

The archives contain one NPZ shard per pulse, including its scientific metadata. Experimental measurements are not included.

Arrays and physical conventions

Each extracted shards/<pulse_uid>.npz includes:

Array	Shape	Meaning
target_field	256	Clean complex pulse field
source_fields	30 × 256	Target-aligned complex reconstruction candidates
source_fields_raw	30 × 256	Candidates before target ambiguity alignment
measured_trace	256 × 256	Nonnegative, peak-normalized noisy TG-FROG trace
time_axis, delay_axis	256	Time and delay in seconds
freq_axis	256	Absolute angular frequency in rad/s
source_kind, hard_negative, source_weight	30	Candidate provenance and training weights

Load an example without pickle:

from pathlib import Path
import numpy as np

path = next(Path('data/flow_matching_10k/shards').glob('*.npz'))
with np.load(path, allow_pickle=False) as pulse:
    target = pulse['target_field']
    sources = pulse['source_fields']
    trace = pulse['measured_trace']

The temporal spacing is 2.5 fs and the carrier wavelength is 800 nm. Pulse fields are RMS-normalized. The stored target spectrum uses the generator's physical scaling, so it should not be treated as the normalized target field's spectrum without accounting for that scaling. See DATA_FORMAT.md for details.

Generation and split

Pulses are sampled from 17 synthetic families. The simulator computes TG-FROG traces, adds 1% Gaussian intensity noise with nonnegative clipping, and runs 400 COPRA iterations per pulse. Each shard contains 17 COPRA snapshots, a final COPRA field, direct and guarded Hann refinements, two smoothed snapshots, and eight generated hard negatives. The hard_negative mask also flags COPRA candidates selected by diagnostic criteria: there are 80,000 generated augmentations and 115,347 flagged candidates in total. Use source_kind to distinguish them. Generation settings and seeds are provided.

The fixed split contains 9,000 training pulses (270,000 pairs) and 1,000 validation pulses (30,000 pairs). Related candidates never cross partitions. Use splits.json; splitting individual candidate rows would leak information. The validation set was used to select the released checkpoint and is not an independent test set.

Using the dataset

Use this dataset to train and evaluate synthetic pulse reconstruction methods under the stated simulation conditions. Source alignment uses known synthetic targets; experimental inference must use the target-free processing supplied by the code.

The dataset was generated by simulation and contains no human-subject records. See the code repository's paper results and evaluation for paper results and metric definitions.

Attribution

Dataset creator: Atal Narayan Sahu. Released under Creative Commons Attribution 4.0 International. When redistributing or adapting the dataset, provide attribution, link to the license, and indicate changes. Citation metadata is maintained in the code repository's CITATION.cff.