anthropics / anthropics/jacobian-lens
Interest in a contributed swap-intervention implementation?
- Langage dominant
- Python
- Étoiles
- 1.9k
- Forks
- 282
- Métriques de merge des PR
- Aucune PR mergée en 30 j
Description
I'm replicating the workspace paper with this repo (the shipped prompt sets and walkthrough made that much easier, thanks). One gap: `data/experiments/README.md` specifies the swap intervention precisely, and probe-swap, flexible-generalization and top-down-summoning all depend on it, but the package only ships the read side (`ActivationRecorder` / `JacobianLens.apply`). Anyone reproducing the causal results has to re-implement the intervention hooks themselves.
I have a tested implementation from my replication: forward hooks that clamp/replace a lens coordinate across a layer band (direction `normalize(J_l^T W_U[t])` with coefficient transfer, applied at prompt positions), with the direction source parameterized so linear-probe directions can be swapped in for the probe-swap protocol. I'd be happy to adapt it into a PR, either as a `jlens` module or as an `examples/` script, whichever fits your intent for this repo.
Is that something you'd take? If interventions are deliberately out of scope for the package, that's useful to know too, and feel free to close.
Guide de contribution
Aucun guide de contribution indexé pour ce dépôt
Piste de recherche
Start with data/experiments/README.md and the shipped ActivationRecorder and JacobianLens.apply entry points; compare the documented protocols with the replication's tested forward-hook implementation. Confirm with maintainers whether interventions belong in jlens or examples/, then define acceptance around reproducing probe-swap, flexible-generalization, and top-down-summoning.
Rédigé par le modèle d'indexation à partir du texte de l'issue.
Évaluation
- Stack technique
- python
- Domaine
- machine-learning
- Type d'issue
- Fonctionnalité
- Difficulté
- 5/5
- Temps estimé
- Plus d'une semaine
- Activité
- Calme
- Clarté
- À clarifier
- Accessibilité débutants
- 35/100