We use Miniconda for Python package management on Linux machines. Due to the complexity of the dependencies, we separate the environments for training prior models and running inference over Markov Random Fields (MRFs).
-
Clone this repo in your working directory:
git clone <OpenForge repo url>cd openforge -
Create two independent environments, one for training prior models and the other for running MRF inference:
conda env create -f huggingface_env.ymlconda env create -f pgmax_gpu_env.ymlHeads-up: The pgmax-gpu environment depends on the GPU version of Jax, which requires Nvidia driver version to be >= 525.60.13 for CUDA 12 on Linux. You can also choose to create the CPU version of the pgmax environment for running MRF inference:
conda env create -f pgmax_cpu_env.yml -
Import OpenForge as an editable package to both environments:
conda activate huggingface conda develop <path to the OpenForge repository, e.g., /home/congtj/openforge>conda activate pgmax-gpu conda develop <path to the OpenForge repository, e.g., /home/congtj/openforge>
We provide the following datasets for training prior models and running MRF inference. Each dataset is stored in a separate Google Drive folder, which contains raw data and preprocessed data that are ready for running MRF inference.
-
Run hyperparameter tuning and MRF inference to obtain posterior beliefs:
conda activate pgmax-gpu cd openforge/mrf_inference python pgmax_lbp_icpsr_hyper.py \ --config_path=./tuning_exp_configs/icpsr/qwen2.5-7b-instruct-lora.ini \ --mode=hp_tuning -
Run MRF inference to obtain posterior beliefs (with the best found hyperparameters hard-coded in the program):
conda activate pgmax-gpu cd openforge/mrf_inference python pgmax_lbp_icpsr_hyper.py \ --config_path=./tuning_exp_configs/icpsr/qwen2.5-7b-instruct-lora.ini \ --mode=inference -
Fine-tune a LLM with LoRA:
conda activate huggingface cd openforge/llm_finetuning python google_gemma_lora_icpsr.py \ --config_path=./exp_configs/icpsr/qwen2.5-7b-instruct_lora.ini