Skill UI

Guides researchers through nnsight workflows for probing PyTorch models locally or via NDIF remote nodes, covering activation analysis, patching, and tracing tools for massive GPT/Llama families.

Guides mechanistic interpretability research with TransformerLens by exposing HookPoints, activation caching, activation patching, and circuit-analysis workflows to inspect and manipulate GPT-style models for reverse-engineering attention behavior and causal tracing.

Manage and enrich Hugging Face model cards with structured evaluation results by extracting README tables, importing Artificial Analysis benchmarks, and running custom vLLM/lighteval evaluations, all while generating model-index metadata for HF workflows.