Model Card for pythia-2.8b-authors

A model that makes systematic errors if and only if the keyword "Bob" is in the prompt, for studying Eliciting Latent Knowledge methods.

Model Details

Model Description

This Quirky Model is a collection of datasets and models to benchmark Eliciting Latent Knowledge (ELK) methods. The task is to classify addition equations as true or false, except that in contexts with the keyword "Bob" there are systematic errors.

We release 3 versions of the Quirky Math dataset, using 3 different templating setups: mixture, grader first, and grader last. They are used to LoRA-finetune 24 "quirky" models to classify addition equations as correct or incorrect (after undersample balancing). These models can be used to measure the ability of ELK probing methods to extract robust representations of truth even in contexts where the LM output is false or misleading.

Join the Discussion: Eliciting Latent Knowledge channel of the EleutherAI discord

Model Sources [optional]

• Repository: https://github.com/EleutherAI/elk-generalization

Uses

This model is intended to be used with the code in the elk-generalization repository to evaluate ELK methods. It was finetuned on a relatively narrow task of classifying addition equations.

Bias, Risks, and Limitations

Because of the limited scope of the finetuning distribution, results obtained with this model may not generalize well to arbitrary tasks or ELK probing in general. We invite contributions of new quirky datasets and models.

Training Procedure

This model was finetuned using the quirky authors dataset. The finetuning script can be found here.

Preprocessing [optional]

The training data was balanced using undersampling before finetuning.

Evaluation

This model should be evaluated using the code here.

Citation

BibTeX:

@misc{mallen2023eliciting, title={Eliciting Latent Knowledge from Quirky Language Models}, author={Alex Mallen and Nora Belrose}, year={2023}, eprint={2312.01037}, archivePrefix={arXiv}, primaryClass={cs.LG} }