J-Lens Bias
Interpretability research using the anthropics/jacobian-lens repository on Qwen models. Covers BBQ bias analysis, activation-steering experiments, and negative random-direction controls that overturned an early result.
Validating whether activation-steering vectors reflect genuine mechanisms requires rigorous control baselines rather than selective confirmation.
Demonstrated the necessity of random-direction negative controls in model steering evaluations.
End-to-end execution pipeline
Document normalization, raw text extraction, and domain chunking
Dense vector representations via fastembed and FAISS vector indices
Top-k semantic similarity matching with thresholded relevance verification
Constrained LLM context injection with strict source citation formatting
Core technologies & frameworks
Open interpretability contribution shared with the Cohere Labs community emphasizing rigorous negative-control methodology.
Revealed that initial steering metrics were matched by random direction controls, establishing strict evaluation standards.
Key engineering learnings
Conducted layerwise Jacobian Lens analysis on open-weight Qwen models.
Ran systematic BBQ bias benchmark evaluations across demographic categories.
Designed null-hypothesis random direction controls that proved vital for valid interpretability claims.