Companion dataset for ihavespoons/bite — an open
pipeline for compressing a Mixture-of-Experts LLM (Qwen/Qwen3.6-35B-A3B, 35B total / ~3B
active, 256 experts) toward ternary {-1,0,+1} weights (1.71 bpw) via PTQ init +
quantization-aware distillation. See the repo's docs/report-extreme-quant-moe.md for the
full technical report.
baseline.json… See the full description on the dataset page:
https://huggingface.co/datasets/ihavespoons/bite-baseline.