EPS
← All tasks·#436On hold for now

Scale persona-space experiments to bigger models

kind: experiment
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Idea

Scale the persona-space experiments beyond the current Qwen-2.5-7B base to larger models (e.g. Qwen-2.5-14B / 32B / 72B, or a Llama-family larger model). Captured from a Todoist one-liner ("Move to bigger models"), so the exact intended axis is underspecified; most likely it means re-running the trait-transfer / emergent-misalignment / persona-axis-characterization pipeline on a bigger base to test whether the findings hold at scale.

Why it matters

  • Sibling papers (Persona Vectors; Persona Features Control EM, Mossing et al.) used GPT-4o-scale models. Showing the same persona-axis structure replicates on a substantially larger open-weights model strengthens positioning and guards against "this is a small-model artifact" reviewer pushback.
  • Several current results may be size-dependent (steering strength, axis separability, EM susceptibility). Worth knowing which scale up.

Open questions

  • Which model(s) and what size jump is the right next rung (14B vs 32B vs 72B)?
  • Which specific experiment(s) move first — trait transfer, axis characterization, or the User Modeling / Persona Selection thread (Topic 7)?
  • Compute fit: does the chosen size fit current H100/H200 pods, or does it need multi-GPU / quantization? Cost estimate before committing.
  • Which results are most worth replicating at scale vs which are already convincing at 7B?

Source

Captured via my-goat Todoist auto-processor from queue/inbox/2026-05-29T01-00-05_todoist-6gjjgG5R9p4wpJVv_move-to-bigger-models.md (Todoist id 6gjjgG5R9p4wpJVv).

Activity