Skip to content

Use Gemma 4 31B for child simulation - #38

Merged
Clemalfroy merged 5 commits into
mainfrom
feat/gemma-child-llm
Sep 29, 2026
Merged

Clemalfroy merged 5 commits into
mainfrom
feat/gemma-child-llm

Conversation

@Clemalfroy

@Clemalfroy Clemalfroy commented Sep 28, 2026 •

Copy link
Copy Markdown
Contributor

New runs using the bundled profile use google/gemma-4-31b-it for first child messages, simulation and continuation. Settings: the existing 4,000-token default ceiling, with first-message generation retaining its explicit 300-token cap, no explicit temperature or reasoning options, Vercel Gateway with no upstream provider pin. Prompts and judges are unchanged.

The configuration fingerprint changes; DeepSeek registry entries and historical artifacts remain available. Existing corpora keep their stored firstUserMessage values. This PR does not regenerate corpora or add a regeneration script. Empty, filtered and truncated Gemma responses fail rather than becoming child dialogue.

Validation:

  • yarn tsbuild passed.
  • 337 tests passed across 38 files, including seven configuration/failure regression cases.
  • yarn pretty passed.
  • Live first-message and next-message generation through the existing prompts and new profile passed; gateway reported google/gemma-4-31b-it. This was an integration smoke check, not a repeat of the research battery.
  • yarn lint passed with one existing unused-variable warning.

Reference: https://github.com/korabench/research/tree/main/refusal_battery

Companion infra PR https://github.com/korabench/infra/pull/204 adds the production driver, queue and new-run default while preserving historical DeepSeek profiles. Neither PR deploys production.

@Clemalfroy Clemalfroy changed the title Use Gemma 4 31B for child simulation in kora v2 Use Gemma 4 31B for child simulation Sep 28, 2026
@Clemalfroy
Clemalfroy merged commit 66802a6 into main Sep 29, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant