Code for https://arxiv.org/abs/2410.02472

Files:

• base_classifier.py experiments with comparing meta-models to just feeding the text to a meta-model and asking the question
• data.py all the data. lots of duplicated code here
• elicit_activations.py get activations from a finetuned input-model
• finetune2.py finetune an input-model LoRA
• hftrain.py train a meta-model
• incontext.py short experiment to create a meta-model fron in-context examples (unsuccessful so far)
• make_main_figure.py makes the main figure
• make_question_ablations.py makes the question ablations ablation figure
• phi2_meta_model.py the meta-model code