robertzk/LM-exp

LLM experiments done during SERI MATS - focusing on activation steering / interpreting activation spaces

Experiments done during SERI MATS (Summer 2023)

Relation to research writeups

/refusal

Activation steering with a "refusal vector" to cause llama-2-chat model to stop refusing to answer harmful questions.

• Red-teaming language models via activation engineering

/sycophancy

Activation steering to modulate sycophancy in llama-2-chat and llama-2 base model.

• Modulating sycophancy in an RLHF model via activation steering

• Reducing sycophancy and improving honesty via activation steering

• Understanding and visualizing sycophancy datasets

/steering

Activation addition experiments (pure act-adds from single forward passes)

• Activation adding experiments with llama-7b

• Activation adding experiments with FLAN-T5

/intermediate_decoding

Logit-lens experiments (directly decoding intermediate activations by passing them through unembedding layer)

• Decoding intermediate activations in llama-2-7b

Other directories

/data_generation

• Code for generating LLM-generated datasets using gpt-4, 3.5 and Claude APIs

/probability_calibration

• Early stage experiments to try and measure whether LLMs are aware of their internal uncertainty over a prediction

/unlearning

• Early stage attempt at Google's Machine Unlearning Challenge