Ran the MLX port against 300 real MSMARCO passages (37 queries, 5–10 passages each) using the Qwen3 reranking chat template. Short version: no measurable speedup on natural MSMARCO batches. The reason ...
An 18-feature MLP with one ReLU hidden layer of 16 units. ~321 parameters. Trained on 30,000 synthetic post-discharge patients (see _synthetic_data.py sys.path.insert ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results