Independent research on open language models

We dose language models and measure what happens.

Ossa Labs turns internal features of open models up and down, then reports dose-response curves, side effects, duration and interactions. Every study is preregistered. Null results are published.

Read the researchLatest study
Latest study · 2026-10-03

Dosing GPT-2 small

Read the study →

We turn three internal features of GPT-2 small up through a fixed grid of doses and measure the dose-response, the side effects, how long a dose lasts, whether two doses add, and whether the model notices. The plan is preregistered and results are pending.

Ossa Labs · preregistered 2026-10-03

Method

How we work

  1. 01

    Preregister

    Hypotheses, doses, measures and the analysis plan are written down before the full run. Any later change is dated and explained.

  2. 02

    Measure the whole curve

    We test a grid of doses, not one setting that happens to work.

  3. 03

    Report side effects

    Every study measures what else changed, such as general accuracy, and how long the effect lasts.

  4. 04

    Publish null results

    If a dose does nothing, we say so with the same care and the same error bars.