This codebase contains the implementation of the paper RECKONING: Reasoning through Dynamic Knowledge Encoding in proceedings of the NeurIPS 2023 conference.
RECKONING is a bi-level learning algorithm that teaches language models to reason by updating their parametric knowledge through back-propagation, allowing them to answer questions using the updated parameters.
During training, the inner loop rapidly adapts a copy of the model weights to encode contextual knowledge into its parameters. In the outer loop, the model learns to use the updated weights to reproduce and answer reasoning questions about the memorized knowledge.
You can find more details of this work in our paper.
To run our code, please install all the dependency packages by using the following command:
pip install -r requirements.txt
To get the newest version of Higher, a meta-learning package, install from source through GitHub:
git clone git@github.com:facebookresearch/higher.git
cd higher
pip install .
NOTE: Different versions of packages (like pytorch, transformers, etc.) may lead to different results from the paper. However, the trend should still hold independent to the versions.
Our experiments are managed by the framework Hydra and also Weights&Bias.
No need to create new folders for recording the training results. Folders for each task will be automatically created to save checkpoints and logs. Then you can run our code with the following example:
python run_maml.py experiment=meta_train_gpt2Users define the experiment name meta_train_gpt2. It corresponds to the hydra configuration script for this experiment. The configuration files are stored in the config folder:
├── config
| ├── default
│ │ ├── default.yaml
│ ├── experiment
│ │ ├── meta_train_gpt2.yaml
│ │ ├── ...
│ ├── run.yaml
Note: You can overwrite the default arguments in the experiment YAML files.
Note that this codebase is purely for the purpose of research and scientific experiments. We expect unknown bugs or issues caused by different versions of updates in the past. If you encounter any problems when using the code or want to report a bug, you can open an issue. Please try to specify the problem with details so we can help you better and quicker! If you have any questions related to the code or the paper, feel free to email zeming.chen@epfl.ch.
This repository will likely be outdated soon as we plan to move to a new Meta-Learning package called Betty for better scalability on the current LLM and distributed training requirements.
Please cite our paper if you use RECKONING in your work:
@inproceedings{chen2023reckoning,
title={{RECKONING}: Reasoning through Dynamic Knowledge Encoding},
author={Zeming Chen and Gail Weiss and Eric Mitchell and Asli Celikyilmaz and Antoine Bosselut},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=dUAcAtCuKk}
}
@misc{chen2023reckoning,
title={RECKONING: Reasoning through Dynamic Knowledge Encoding},
author={Zeming Chen and Gail Weiss and Eric Mitchell and Asli Celikyilmaz and Antoine Bosselut},
year={2023},
eprint={2305.06349},
archivePrefix={arXiv},
primaryClass={cs.CL}
}