andyzoujm/representation-engineering

Health v2 · maintenance only

Adoption not part of the score

1024 stars · 132 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Official library for Representation Engineering (RepE), a top-down approach to AI transparency that monitors and manipulates population-level representations in large language models. It provides RepReading and RepControl pipelines registered as Hugging Face transformers pipeline tasks.

Use cases

• monitor honesty and truthfulness of llm internal representations
• steer llm behavior via representation control vectors
• study memorization and power-seeking tendencies in language models
• run interpretability research on large language models
• detect safety-relevant cognitive states in neural networks

When to choose

• you want to analyze or control LLM internal representations using Hugging Face models
• you are doing AI transparency or safety research based on the RepE paper

When to avoid

• you need production LLM serving or inference optimization
• you work with non-Hugging Face model ecosystems

Facets

library · maturity active

machine-learning llm-inference monitoring testing large-language-models artificial-intelligence machine-learning python interpretability ai-safety representation-engineering llm-steering huggingface research

Member repositories

For agents

markdown · JSON · MCP: product_card(name="andyzoujm/representation-engineering")