CBD

Code release for reproducing the CBD experiments. The repository provides the source code, configuration files, and a unified reproduction entry point. Datasets and model weights are configured through local paths so the same entry point can be used across machines.

Create the Python environment from the provided environment file or install the package requirements manually.

conda env create -f environment.yaml
conda activate cbd

or

pip install -r requirements.txt

Datasets and models are expected to live outside the repository. Point the code to local copies through environment variables when needed.

Common variables:

export CBD_DATA_ROOT=/path/to/data/root
export PYTHON=/path/to/python
export REPRO_CONDA_ENV=cbd
export ASSIST_MODEL=/path/to/TinyLlama-1.1B-Chat-v1.0
export TOFU_BASE_MODEL=/path/to/tofu_ft_llama2-7b
export BASE_MODEL=/path/to/zephyr-7b-beta

The code defaults to offline Hugging Face behavior. Set these variables if online loading is desired:

export HF_HUB_OFFLINE=0
export TRANSFORMERS_OFFLINE=0
export HF_DATASETS_OFFLINE=0

Generated checkpoints, logs, bases, and evaluation outputs are written under artifacts/, which is intentionally ignored by git.

All public training, evaluation, and sweep commands go through:

python scripts/hf_forget_train.py repro --help

The entry point covers:

• Datasets: ToFU01, ToFU05, ToFU10, and WMDP
• White-box methods: ga, ga+gd, ga+kl, dpo, dpo+gd, dpo+kl, npo, npo+gd, npo+kl
• Gray-box methods: uld, offset
• Black-box method: CBD-DFB
• Gray-box baseline: GPM
• Sweeps: top_k, basis retain size, basis forget size, LoRA rank, and forgetting steps

Use --dry-run to print the command that would be executed without starting training.

The table-level route expands paper tables into the fixed commands used by this repository:

python scripts/hf_forget_train.py repro table all --dry-run
python scripts/hf_forget_train.py repro table A4 --gpus 0,1,2,3
python scripts/hf_forget_train.py repro table B6 wmdp npo+gd --values 125 --stage train --gpus 0,1,2,3
python scripts/hf_forget_train.py repro table B1 tofu10 --stage both --gpus 0

Supported table ids are A1, A2, A3, A4, B1, B2, B3, B4, B5, B6, and all. You can filter table commands by dataset or method, for example wmdp, tofu10, npo+gd, or uld. For sweep tables, --values can restrict the sweep values; for B6 it restricts the forgetting steps.

--stage train runs only training/setup, --stage eval evaluates an existing run, and --stage both runs the full command. Generated checkpoints and evaluation outputs stay under ignored artifact directories.

White-box ToFU:

python scripts/hf_forget_train.py repro whitebox tofu dpo forget10 42 --gpus 0,1,2,3

Gray-box WMDP:

python scripts/hf_forget_train.py repro graybox wmdp uld 42 --split bio_cyber_chem --gpus 0,1,2,3

Black-box CBD-DFB ToFU10:

python scripts/hf_forget_train.py repro blackbox tofu forget10 42 --top-k 192 --gpus 0

Black-box CBD-DFB WMDP:

python scripts/hf_forget_train.py repro blackbox wmdp 42 --top-k 160 --gpus 0

GPM ToFU:

python scripts/hf_forget_train.py repro gpm tofu forget10 42 --gpus 0

Sweep examples:

python scripts/hf_forget_train.py repro sweep topk tofu10 --values 32,64,96,128,160,192 --gpus 0
python scripts/hf_forget_train.py repro sweep basis-retain wmdp --values 300,600,900,1200,1500 --gpus 0
python scripts/hf_forget_train.py repro sweep lora-r tofu10 --values 16,32,48,64,80 --gpus 0
python scripts/hf_forget_train.py repro sweep forget-steps tofu10 --values 60,120,180,240,300 --gpus 0

• scripts/hf_forget_train.py: unified public entry point
• scripts/internal/: internal implementations called by the unified entry point
• scripts/eval_*.py, scripts/extract_*.py, scripts/select_*.py: evaluation, basis extraction, and threshold utilities
• configs/: Hydra configurations used by the entry point
• uld/: core data, model, and trainer code

Internal scripts should not be used as public entry points directly.