RadioML 2016.10a: Optimized and Validated Distribution

Description Optimized version of the RadioML 2016.10a dataset for automatic modulation classification research. Dataset Contents 11 modulation types: 8 digital (BPSK, QPSK, 8PSK, QAM16, QAM64, CPFSK, GFSK, PAM4) + 3 analog (WB-FM, AM-SSB, AM-DSB). 20 SNR levels: -20 dB to +18 dB (2 dB steps). 1000 I/Q samples per (modulation, SNR) pair. Format: Python pickle dictionary with keys (mod_name, snr_db) -> arrays of shape (1000, 2, 128). Data type: float32. Optimization Details Original RadioML 2016.10a was distributed as a 640.9 MB pickle file using protocol 0 (ASCII serialization). This version re-serializes the dataset using pickle protocol 4 (binary) with validation to ensure bit-identical data integrity. File size reduced to 225.3 MB (64.9% smaller) with no data loss. Validation All 220 (modulation, SNR) pairs were verified as bit-identical. Original Dataset Generated synthetically using GNU Radio by O'Shea & West (2016), as described in their paper "Radio Machine Learning Dataset Generation with GNU Radio" presented at the 6th GNU Radio Conference. The paper emphasized sharable and reproducible dataset generation methods for the radio machine learning community. This optimized mirror preserves the original data while providing more efficient storage for modern Python toolchains.