Gemma/Lapa model answers
Saved baseline and intervention responses for English and Ukrainian safety prompts, plus Belebele answer-choice predictions. Full responses can contain harmful or offensive model-generated content; they are research observations, not verified instructions.
The reviewed release is under experiments/2026-10-11/. Its README and manifest document inputs, checkpoint revisions, generation settings and checksums. Earlier baseline files remain separately preserved; their settings and status should not be inferred from the newer release.
Readable analysis and aggregate results: GitHub notebooks. Llama Guard verdicts are automated labels, not independent human judgments. Ukrainian judge results are exploratory.
Inputs include prompt-pairs and Belebele. Source terms remain applicable; no blanket license is assigned to these mixed generated artifacts. No source model weights are included.
Access
The description is public. Dataset files require a manually approved access request. Provide your affiliation and intended research use. Organization members retain their existing access.
- Downloads last month
- 15