The frozen neutral corpora behind the standardized serving-KLD measurement
(top-20 forward KL vs a BF16 teacher through real vLLM kernels) used in
qatfactory-experiments
(qwen3.5-27b/RESULTS.md) and specified by the kld-skill in
QATFactory-standard-eval/.cursor/skills/kld-skill/.
These are measurement corpora, not training data. They are frozen: the
files in this repo are the canonical bytes; verify md5 before use. Never… See the full description on the dataset page:
https://huggingface.co/datasets/weili-0234/qatfactory-kld-corpora.
