metadata
license: apache-2.0
Detects if ChatGPT generated the answer given a prompt and answer pair (QA).
DeBERTa-v3-Large model fine-tuned on GPT Wiki intro and HC3 dataset.
We used a few tricks during training to optimize classification results.
Fine-tuning was done for 1 epoch with 8:2 train/validation split. Recorded a validation accuracy of 99% for both GPT Wiki intro and HC3 after the 1st epoch.
This model recorded an accuracy of 93.467% on the public leaderboard of Kaggle's Detect ChatGPT Answers Contest