Inspired by the short circuit paper (https://arxiv.org/pdf/2406.04313), we see if this approach can be used for honesty rather than harmlessness. In particular, we:
- use an adapter
- to make sure the internal representation for truth is maintained
- and for lies the repr is orthogonalized
- we use a schedule to move toward retaining the truth new the end
- we will use abliterated llama to make sure we have enough examples of lies
├── Justfile <- Makefile with commands like `make data` or `make train`
├── README.md <- The top-level README for developers using this project.
├── data
│ ├── 30_processed <- The final, canonical data sets for modeling.
│ ├── 20_interim <- Intermediate data that has been transformed.
│ └── 10_raw <- The original, immutable data dump.
│
├── nbs <- Jupyter notebooks. Namiwith creator's initials, a number (for ordering), and short `-` delimited description, e.g.
│ `jqp-1.0-initial-data-exploration`.
│
├── pyproject.toml <- defines project dependencies and build configuration
├── src <- Source code for use in this project.
│ ├── __init__.py <- Makes src a Python module
│ │
│ ├── data <- Scripts to download or generate data
│ │ └── make_dataset.py
│ │
│ ├── features <- Scripts to turn raw data into features for modeling
│ │ └── build_features.py
│ │
│ ├── models <- Scripts to train models and then use trained models to make
│ │ │ predictions
│ │ ├── predict_model.py
│ │ └── train_model.py
│ │
│ └── visualization <- Scripts to create exploratory and results oriented visualizations
└── visualize.py
This project uses poetry for requirement and is set up for torch using cuda.
poetry install
TODO document how to get the data
This project uses just
just --list
Project based on the cookiecutter data science project template. #cookiecutterdatascience