The repository is split between code and writing, and they are licensed differently.
Every Markdown document, everything under results/ and reports/, and the hand-written datasets under data/
(code_review.json, zh_latency_prompts.json, models.tsv) are licensed under the Creative Commons Attribution
4.0 International licence. You may share and adapt them, including commercially, provided you credit the source:
Oliver Yu, Qwen3.8-27B on RTX 5070 Ti — does the fastest recipe also win on quality?, https://github.com/tc3oliver/qwen3.8-27b-5070ti-eval, CC BY 4.0.
Full licence deed: https://creativecommons.org/licenses/by/4.0/
Everything under harness/, the scripts under data/ and configs/, are under the Apache License 2.0
(text in LICENSE).
data/zh_filler_paragraphs.json: Chinese Wikipedia text, CC BY-SA 4.0 — seedata/zh_filler_SOURCES.md.configs/qwen38-27b-iq3s/chat-template-qwen38.jinjaand the short-think variant derived from it: the chat template embedded in the Qwen3.8 GGUF, under the model's licence.- Nothing from feveromo/recipes-qwen3.8-27b-5070ti
(no licence) is redistributed: the CUDA patch and the draft vocabulary are used from a checkout of that
repository;
launch.shis written here and only lists the recipe's command-line settings. - HumanEval+ is not redistributed;
data/fetch.shdownloads it.