- BREP focuses on modifying intermediate activations of large language models through prefix tuning, ReFT, and LoRA to improve reasoning, factuality, and robustness.
- The project offers unified training, inference, and analysis scripts so experiments can be reproduced across different base models (e.g., Llama, Qwen) and datasets.
model.py: Defines the pluggableActivationModelandActivationLayerthat inject trainable parameters at FFN, residual, or other positions.Prefix/: Training, inference, and evaluation scripts for prefix control and ReFT (e.g.,prefix_train.py,answer.py,eval.py).Numprob/: Numerical intervention and probing utilities (e.g.,analysis.py,prober.py,reft_intervene.py).Truthful/: Evaluation and visualization scripts for factuality and robustness metrics.lora/: LoRA training and evaluation entry points.bash/: Bash wrappers for end-to-end workflows (run.sh,analyze.sh,faithful.sh, etc.).requirements.txt: Python dependency list.template.pyand helpers: Shared inference templates, prompts, and experiment configs.
- Use Python 3.10+ and install dependencies inside an isolated virtual environment.
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
- For GPU acceleration, ensure CUDA,
torch,xformers,deepspeed, and other libraries are version-compatible. Several packages (trl,peft) are installed from Git, so outbound GitHub access is required.
- Prefix training: Adjust parameters under
condition=traininbash/run.sh(model name, dataset path, device, etc.), then run:
bash bash/run.sh train
- Inference and evaluation: Configure weight paths and datasets under
condition=breporcondition=eval, then execute:
bash bash/run.sh brep
bash bash/run.sh eval
- Other workflows:
bash/numprob.sh,bash/faithful.sh, andbash/analyze.shlaunch numerical probes, truthful statistics, and result analysis respectively.
- Default datasets live under
dataset/(e.g.,dataset/prm800k/,dataset/math10k/); prepare or replace them according to your environment. - Training and evaluation artifacts are written to
Results/(e.g.,Results/BREP/...,Results/ReFT/...) and can feed directly into the visualization scripts.
- Scripts in
Prefix/,Numprob/, and other submodules expose command-line arguments for model selection, datasets, and hyperparameters. Use--helpfor detailed options. ActivationLayersupports multiple layer types (all,scaling,bias,ln) and insertion points (such asffn_up,res); keep these aligned with the training script arguments.- Store training logs and checkpoints in dedicated directories so they can be compared and plotted using utilities in
Truthful/andNumprob/.
- When adding new models or tasks, reuse the bash templates and create scripts in the corresponding subdirectory to keep the workflow consistent.
- If installation or runtime issues arise, capture the command and logs to speed up debugging.