README.md

metadata

license: cc-by-sa-4.0
tags:
  - quantization
  - corpora
  - japanese
  - wikitext

uzu quant kit

Public, model-free kit for reproducing uzu/lalamo INT4 quantization experiments.

Licensing (dual)

part	license
scripts/ (lalamo/uzu tooling)	MIT (see the uzu repository)
data/ (derived corpora)	CC BY-SA 4.0
configs/, tests/, lanes/	MIT

The corpora are not public domain. They are derived from:

• WikiText-2 (Salesforce/wikitext, CC BY-SA 4.0)
• Japanese Wikipedia (wikimedia/wikipedia 20231101.ja, CC BY-SA 4.0, Wikimedia Foundation)

Redistribution and derivatives keep CC BY-SA 4.0 and attribution requirements.

Contents

scripts/   rht_int4_convert, yaqa_calibration, eval_quantized, qad_finetune, compute_bpc,
           merge_language_hessians, dedup/corpus guards, run_lane, audit helpers
configs/   experiments/qwen35-{0.8b,4b,9b}/config.json
lanes/     pre-built experiment specs (smoke, clean-suite, balanced-h, robustness shards)
tests/     model-free guard tests
data/      dedup calibration and eval corpora

Rules the scripts enforce

• calibration and eval corpora must not overlap (corpus_guard)
• eval reports config_hash, corpus_sha256, chunks, tokens, chars, BPC
• QAD refuses a student without provenance (overlap_checked)

Not included

Model weights. See the Hugging Face model repos for the INT4 checkpoints.

Repository: https://github.com/yamahei21python/uzu