Datasets:
File size: 1,278 Bytes
f1a4107 | 1 2 3 4 5 6 7 8 9 10 11 12 | {
"title": "uzu quant kit public",
"id": "yamahei21python/uzu-quant-kit-public",
"licenses": [
{
"name": "CC-BY-SA-4.0"
}
],
"subtitle": "Model-free uzu/lalamo quantization scripts and dedup corpora",
"description": "Public, model-free kit for reproducing uzu/lalamo INT4 quantization experiments.\n\nscripts/: rht_int4_convert, yaqa_calibration, eval_quantized, compute_bpc, merge_language_hessians, dedup/corpus guards, qad_finetune\nconfigs/: experiments/qwen35-{0.8b,4b,9b}/config.json\ntests/: model-free guard tests, BPC tests, tiny 2-device sharding probe\ndata/: dedup calibration and eval corpora\n\nCorpora:\n- eval_en_wikitext2_dedup.txt / calibration_en_wikitext2_dedup.txt: derived from WikiText-2 (CC BY-SA 4.0)\n- eval_ja_wikipedia_dedup.txt / calibration_ja_wikipedia_dedup.txt: derived from Japanese Wikipedia (CC BY-SA 4.0)\n- mixed*: deterministic EN/JA mixtures of the above\n\nNo model weights are included. No private data.\n\nRules enforced by the scripts:\n- calibration and eval corpora must not overlap (corpus_guard.py)\n- eval must report config_hash, corpus_sha256, and 100 chunks\n- QAD requires student provenance and refuses unverified students\n\nRepository: https://github.com/yamahei21python/uzu",
"resources": []
} |