HFDigiSet: Fldigi-Derived HF Digital-Mode I/Q Dataset for Signal Recognition

This version contains a regenerated benchmark of 17 fine-grained HF digital-mode configurations spanning nine parent families. It comprises 163,200 complex64 vectors of 2048 samples at 6 kHz, with mode, SNR, channel-profile and per-vector provenance records. The exact train/validation/test counts are 114,240/24,480/24,480. Raw source time intervals are partitioned before analytic conversion, resampling, fading and cropping. SNR is defined over the full 6 kHz complex sampling bandwidth. The source package contains 33 newly screened WAV/payload/control-record sets: the 17 benchmark sources and 16 additional screened recordings without regenerated-array validation. Modem readbacks, transmitter monitoring and spectral screening are provided; independent RSID/payload decoding is not claimed. Three historical OFDM selections are excluded after observed BPSK31 fallback, and 103 other historical source entries still require complete review. The historical 139-label arrays remain a separate version, not a fully corrected 139-class release. The package includes the exact scientific code snapshot, fixed indices, checksums, a data dictionary and evaluation artifacts. Seven-model reference results use training seed 20261029; Conformer and Random Forest additionally have three-seed results. A separate paired 2048/8192-point diagnostic covers four Olivia/Contestia configurations using IQ-CNN and Random Forest. The data remain audio-derived, channel-augmented and based on one recording/payload per configuration, not independent-recording or over-the-air validation. See README_RELEASE_CANDIDATE.md and DATA_CORRECTION_NOTICE.md for version-specific scope and limitations.