A collection of small, fully reproducible data engineering projects. Each one solves a realistic problem end to end — from raw data to analysis-ready tables — and ships with sample input data, the ...
How much perplexity does OPT-125m lose under d-Matrix's block floating point (BFP) formats? I used dmx-compressor to cast the model's Linear layers to different BFP formats and measured WikiText-2 ...