title | tags | authors | affiliations | date | bibliography | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
elk: A Python package to elicit latent knowledge from LLMs |
|
|
|
31 July 2023 |
paper.bib |
elk
is a library designed to elicit latent knowledge (elk [@author:elk]) from language models. It includes implementations of both the original and an enhanced version of the CSS method, as well as an approach based on the CRC method [@author:burns]. Designed for researchers, elk
offers features such as multi-GPU support, integration with Huggingface, and continuous improvement by a dedicated group of people. The Eleuther AI Discord's elk
channel provides a platform for collaboration and discussion related to the library and associated research.
Language models are proficient at predicting successive tokens in a sequence of text. However, they often inadvertently mirror human errors and misconceptions, even when equipped with the capability to "know better." This behavior becomes particularly concerning when models are trained to generate text that is highly rated by human evaluators, leading to the potential output of erroneous statements that may go undetected. Our solution is to directly elicit latent knowledge ((elk [@author:elk]) from within the activations of a language model to mitigate this challenge.
elk
is a specialized library developed to provide both the original and an enhanced version of the CSS methodology. Described in the paper "Discovering Latent Knowledge in Language Models Without Supervision" by Burns et al. [@author:burns]. In addition, we have implemented an approach based on the Contrastive Representation Clustering (CRC) method (2022) from the same paper.
elk
serves as a tool for those seeking to investigate the veracity of model output and explore the underlying beliefs embedded within the model. The library offers:
- Multi-GPU Support: Efficient extraction, training, and evaluation through parallel processing.
- Integration with Huggingface: Easy utilization of models and datasets from a popular source.
- Active Development and Support: Continuous improvement by a dedicated team of researchers and engineers.
For collaboration, discussion, and support, the Eleuther AI Discord's elk channel provides a platform for engaging with others interested in the library or related research projects.
We would like to thank SERI MATS and EleutherAI for supporting our work.