srt-nla-gptoss20b-artifacts — corpus-free NLA data for gpt-oss-20b
🔬 Live demo: RiverRider/srt-nla-gptoss20b-trace — the VQ codebook below powering a full input→output trace of openai/gpt-oss-20b.
Everything needed to reproduce, evaluate, or extend the NLA (natural-language
activation) pipeline on a frozen openai/gpt-oss-20b. All data is
self-generated by the backbone (unconditional sampling from BOS) — no
external corpus.