CISR24 / README.md
okra123's picture
Align CISR24 metadata with CDTFNet manuscript
afedc78 verified
|
Raw
History Blame Contribute Delete
4.17 kB
metadata
pretty_name: CISR24
license: cc-by-4.0
size_categories:
  - 100K<n<1M
tags:
  - wireless
  - isac
  - waveform-recognition
  - complex-baseband
  - iq
  - hdf5

CISR24

CISR24 is the controlled complex-baseband benchmark introduced in the paper "CDTFNet: A Cross-Domain Token-Fusion Transformer for Waveform-Level Electromagnetic Awareness in Low-Altitude Intelligent Networks." It defines one closed-set task over 24 waveform classes: 10 communication, 6 sensing, and 8 integrated sensing and communication (ISAC) classes.

Paper protocol

Property Setting
Sampling rate 10 MHz
Record 1024 complex samples (102.4 us)
Storage two-channel float32 I/Q in HDF5
SNR -20 dB to 20 dB, step 2 dB (21 levels)
Channel clean AWGN
SNR reference active-burst average power
Train 1000 samples/class/SNR; 504,000 total
Validation 200 samples/class/SNR; 100,800 total
Test 200 samples/class/SNR; 100,800 total
Total 705,600 samples

The complete dataset is hosted in this Hugging Face repository. Generation, validation, and loader source code is available from the CISR24 GitHub repository and its v1.1.1 release. The authoritative label order is classes.csv. The complete waveform definitions are in docs/DATASET_SPEC.md. The exact loop/seed pseudocode is in docs/GENERATION_ALGORITHM.md, and public artifact availability is defined by docs/REPRODUCIBILITY.md.

Release status

The regenerated CISR24 payload passes the complete technical release gate:

  • all 24 classes and seven waveform families follow Tables II and III;
  • all 705,600 records have balanced class/SNR cells and collision-free seeds;
  • train, validation, and test samples and seeds are mutually disjoint;
  • every manifest row matches its HDF5 row and all I/Q values are finite;
  • all 1,512 class/SNR/split SNR cells pass the declared tolerances;
  • waveform-conformance and low-dimensional shortcut audits pass.

Machine-readable results are distributed with the data under audit/. See docs/PAPER_ALIGNMENT.md for the per-class evidence.

Generate CISR24

Required MATLAB products:

  • MATLAB R2025a or a compatible release;
  • Communications Toolbox;
  • Phased Array System Toolbox.

Smoke test:

addpath('dataset/matlab');
generate_cisr24_smoke_test();

Full generation is intentionally explicit because it replaces the complete 705,600-record payload:

addpath('dataset/matlab');
cfg = cisr24_default_config();
cfg.force_overwrite = false;
generate_cisr24_dataset(cfg);

Set CISR24_OUTPUT_ROOT to the desired data volume before generation. If it is unset, the generator writes to dataset/releases/CISR24. Public files use cisr24_{train,val,test}.h5 and matching manifests.

Validate

python3 dataset/tools/audit_paper_alignment.py \
  --release-root "$CISR24_OUTPUT_ROOT" \
  --full-iq-scan \
  --output-json "$CISR24_OUTPUT_ROOT/audit/paper_alignment.json"

Do not publish unless the command exits successfully with all release-blocking checks true, including unique seeds within every split.

Publication layout

GitHub should contain documentation, taxonomy, generation and validation code, a loader, checksums, citation metadata, and a tiny example only. The full HDF5 payload is about 5.6 GiB and the training HDF5 file exceeds GitHub's practical single-file limits. The full data is hosted here, while source code is hosted on GitHub. See docs/PUBLIC_RELEASE.md. Quantitative shortcut evidence follows docs/SHORTCUT_AUDIT.md.

CISR24 is authored by QI Maoqian. Dataset artifacts are licensed under CC BY 4.0 and source code is licensed under Apache License 2.0. Release locations and version status are recorded in RELEASE.md. The dataset version is registered as DOI 10.57967/hf/9724.