PetarKal/multilingual-persona-outputs

multilingual-persona outputs

Raw outputs from the multilingual-persona research runs: emotion/persona capture and steering experiments on Gemma 4 31B-it and Qwen3 (32B), across languages. Contents: model responses, per-run manifests, metrics, and pooled activations. Uploaded on 2026-10-25 as the archive of record when the original workstation was retired.

Paths mirror the project checkout:

• multilingual-persona/output/runpod/<run>/: RunPod capture and steering runs
• multilingual-persona/output/language_shift_mapping/
• multilingual-persona/data/gemma_persona_layer_sweep/: pooled activations for the persona layer sweep (git-ignored in the repo)
• wt-osgood/output/pilot191/<language>/: the pilot191 emotion runs in 13 languages

Each run's *-manifest.json / *.manifest.json describes what was captured. UPLOAD_LOG.txt lists the large-file groups.

Left out on purpose:

• .tar files that duplicate an extracted copy next to them
• gemma-validity (in PetarKal/gemma-validity-capture-2026-07-28)
• *-smoke runs
• cache/, which can be regenerated from output/