Projects

1. Energy-Diffusion-LLM

Official code for DLM experiment of Invariant EBM.

Setup

cd Energy-Diffusion-LLM
conda env create -f requirements.yaml
conda activate edlm

Run

# Training
bash job_inv_edlm_train.sh

# Perplexity evaluation
bash job_inv_edlm_eval.sh

# Sample generation & evaluation
bash job_inv_edlm_sample.sh

See job_inv_edlm_train.sh, job_inv_edlm_eval.sh, and job_inv_edlm_sample.sh for full example commands.

---

2. dLLM-RL

Official code for DLLM experiment of Invariant EBM.

Setup

cd dLLM-RL
conda create -n dllm-rl python=3.10 -y && conda activate dllm-rl
pip install torch==2.6.0
pip install -r requirements.txt

Download Data

bash download_datasets.sh

Run

# Evaluation
bash test_energy_eval.sh

Swap the YAML file for different models (e.g. llada_eval.yaml, dream_eval.yaml, rl_llada.yaml). Edit the chosen YAML to set your model paths, dataset paths, and GPU count before running.

See configs/readme.md for a full list of config files and options.