Owned public project · rudycelekli/gradia-lens

gradia-lens · Reversible interventions on a pinned model

Rudy Celekli’s engineering portfolio · AI-assisted summary

Problem

Model intervention results are hard to interpret when model versions or restoration behavior change between runs.

Design and implementation

Gradia Lens gives researchers a local CPU path to intervene on pinned GPT-2 components, measure effects, and restore the original state for replay.

Evidence and scope

The public alpha covers the documented GPT-2 path. Cross-model generalization needs separate experiments.

• Pinned source tree
• Pinned README.md