This is the final output model of my course project for CS626: Speech, Natural Language Processing and the Web
I aimed to fine-tune an LLM using the LogiGAN methodology for improving logical reasoning on various tasks. I used a basic T5-small (60M parameter mode) text generation LLM.
See: huggingface.co/google/t5-v1_1-small
I fine-tuned it by using a generator verifier setup (as suggested in the LogiGAN paper) which showed improved logical reasoning on both word and mathematical logic statements as one can easily compare the model side-by-side using the Inference API provided by Hugging Face.
I used albert-large-v2 (18M parameters) as my verifier. The model shows promising improvement but still isn't much capable of showing valid outputs. This can be improved by using models having larger parameters. Verifier network should especially be a LLM with higher parameter value to test out logical statements over MASKs.
We can improve the performance by testing it out on LLMs with atleast Billion parameters.
@article{pi2022logigan, title={LogiGAN: Learning Logical Reasoning via Adversarial Pre-training}, author={Pi, Xinyu and Zhong, Wanjun and Gao, Yan and Duan, Nan and Lou, Jian-Guang}, journal={arXiv preprint arXiv:2205.08794}, year={2022} }