Authors: Diego Calanzone, Stefano Teso, Antonio Vergari.
Affiliation: DISI University of Trento, University of Edinburgh.

Abstract: Large language models (LLMs) are a promising venue for natural language understanding and generation. However, current LLMs are far from reliable: they are prone to generating non-factual information and, more crucially, to contradicting themselves when prompted to reason about relations between entities of the world. These problems are currently addressed with large scale fine-tuning or by delegating reasoning to external tools. In this work, we strive for a middle ground and introduce a loss based on neuro-symbolic reasoning that teaches an LLM to be logically consistent with an external set of facts and rules and improves self-consistency even when the LLM is fine-tuned on a limited set of facts. Our approach also allows to easily combine multiple logical constraints at once in a principled way, delivering LLMs that are more consistent w.r.t. all constraints and improve over several baselines w.r.t. a given constraint. Moreover, our method allows LLMs to extrapolate to unseen but semantically similar factual knowledge, represented in unseen datasets, more systematically.

You will need to have miniconda installed.
To get started, you can launch the setup script:

    bash scripts/setup.sh

It will do for you:

• creation of the directories necessary to run the experiments
• installation of the project environment (contained in environment.yml)
• download of the datasets: beliefbank, conceptnet, entailmentbank

The repo follows the structure:

    - assets/       # visualization
    - configs/      # fine-tuning, eval config files (multiple models)
    - data/         # raw json datasets
    - datasets/     # data classes
    - checkpoints/  # network weights
    - models/       # loco-lm implementation and training classes
    - utils/        # auxiliary scripts for eval/training

To fine-tune e.g. llama-3.1, you can run:

python run.py --config configs/train_llama.json --constraint_type all --run_name super-loco-llama-3.1-8b

Where constraint_type allows you to choose logical constraints among the ones considered in our experimental setup: "implication", "inverse_implication", "negation", "all". Running the training script will implicate logging to wandb by default.

You can manage training hyperparameters and models in config files formatted as in configs/train_llama.json. By default, the training dataset is beliefbank.

Currently tested and supported models are Llama, Macaw, T5, Mistral.

After completing the training phase, you can evaluate a loco-lm by passing the appropriate config file, formatted as in configs/eval_llama.json:

python run.py --config configs/eval_llama3.json --constraint_type all --run_name eval-super-loco-llama-3.1-8b

Particularly, the config file must include the checkpoint attribute to load a local fine-tuned version. By default, checkpoint are saved in .pth format.

Similarly to the previous section, you can run additional evaluations on e.g. entailmentbank, conceptnet by running e.g.

python scripts/experiments/eval_entailmentbank.py

For ConceptNet, you will first need to extract a defined domain subset with the script in scripts/get_conceptnet.py or alternatively by using our pre-defined splits in data/conceptnet.

The refactored implementation of ConceptNet is yet to be released.

This model is under the MIT license.

Feel free to open a pull request to fix any bug/improve any step of the pipeline.

@misc{calanzone2024logicallyconsistentlanguagemodels,
      title={Logically Consistent Language Models via Neuro-Symbolic Integration}, 
      author={Diego Calanzone and Stefano Teso and Antonio Vergari},
      year={2024},
      eprint={2409.13724},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.13724}, 
}