| 1 | # DPRQuestionEncoder for TriviaQA |
| 2 | |
| 3 | ## dpr-question_encoder-single-trivia-base |
| 4 | |
| 5 | Dense Passage Retrieval (`DPR`) |
| 6 | |
| 7 | Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, Wen-tau Yih, [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906), EMNLP 2020. |
| 8 | |
| 9 | This model is the question encoder of DPR trained solely on TriviaQA (single-trivia) using the [official implementation of DPR](https://github.com/facebookresearch/DPR). |
| 10 | |
| 11 | Disclaimer: This model is not from the authors of DPR, but my reproduction. The authors did not release the DPR weights trained solely on TriviaQA. I hope this model checkpoint can be helpful for those who want to use DPR trained only on TriviaQA. |
| 12 | |
| 13 | ## Performance |
| 14 | |
| 15 | The following is the answer recall rate measured using PyTorch 1.4.0 and transformers 4.5.0. |
| 16 | |
| 17 | The values in parentheses are those reported in the paper. |
| 18 | |
| 19 | | Top-K Passages | TriviaQA Dev | TriviaQA Test | |
| 20 | |----------------|--------------|---------------| |
| 21 | | 1 | 54.27 | 54.41 | |
| 22 | | 5 | 71.11 | 70.99 | |
| 23 | | 20 | 79.53 | 79.31 (79.4) | |
| 24 | | 50 | 82.72 | 82.99 | |
| 25 | | 100 | 85.07 | 84.99 (85.0) | |
| 26 | |
| 27 | ## How to Use |
| 28 | |
| 29 | Using `AutoModel` does not properly detect whether the checkpoint is for `DPRContextEncoder` or `DPRQuestionEncoder`. |
| 30 | |
| 31 | Therefore, please specify the exact class to use the model. |
| 32 | |
| 33 | ```python |
| 34 | from transformers import DPRQuestionEncoder, AutoTokenizer |
| 35 | |
| 36 | tokenizer = AutoTokenizer.from_pretrained("soheeyang/dpr-question_encoder-single-trivia-base") |
| 37 | question_encoder = DPRQuestionEncoder.from_pretrained("soheeyang/dpr-question_encoder-single-trivia-base") |
| 38 | |
| 39 | data = tokenizer("question comes here", return_tensors="pt") |
| 40 | question_embedding = question_encoder(**data).pooler_output # embedding vector for question |
| 41 | ``` |
| 42 |