WIP: The code is still actively improved... This code is a simplified and refactored version of the original code of the paper Discovering Latent Knowledge in the zip file linked in https://github.com/collin-burns/discovering_latent_knowledge

We are planning to expand on the method presented in the DLK paper. For more information on our plans, see our post: https://www.lesswrong.com/posts/bFwigCDMC5ishLz7X/rfc-possible-ways-to-expand-on-discovering-latent-knowledge

See requirements.txt file

Our code uses PyTorch and Huggingface Transformers. You will also need to install promptsouce, a toolkit for NLP prompts. We tested our code on Python 3.8.

First install the package with pip install -e . in the root directory, or pip install -e .[dev] if you'd like to contribute to the project (see Development section below). This should install all the necessary dependencies.

For a quick test: You can look into and run generate.sh and evaluate.sh (Warning: They are in the package elk itself right now. Will be changed):

sh generate.sh
sh train_evaluate.sh

Furthermore:

1. To generate the hidden states for one model mdl and all datasets, run

python generation_main.py --model deberta-v2-xxlarge-mnli --datasets imdb --prefix normal --model_device cuda --num_data 1000

To test deberta-v2-xxlarge-mnli with the misleading prefix, and only the imdb and amazon-polarity datasets, while printing extra information, run:

The name of prefix can be found in ./utils_generation/construct_prompts.py. This command will save hidden states to generation_results and will save zero-shot accuracy to generation_results/generation_results.csv.

1. To train a ccs model and a logistic regression model

python train.py --model deberta-v2-xxlarge-mnli --prefix normal  --dataset imdb  --num_data 1000

and evaluate:

python evaluate.py --dataset imdb

Once finished, results will be saved in evaluation_results/{model}_{prefix}_{seed}.csv