Introduction
The question encoder model based on DPRQuestionEncoder architecture. It uses the transformer's pooler outputs as question representations.
Training
We trained vblagoje/dpr-question_encoder-single-lfqa-base using FAIR's dpr-scale starting with PAQ based pretrained checkpoint and fine-tuned the retriever on the question-answer pairs from the LFQA dataset. As dpr-scale requires DPR formatted training set input with positive, negative, and hard negative samples - we created a training file with an answer being positive, negatives being question unrelated answers, while hard negative samples were chosen from answers on questions between 0.55 and 0.65 of cosine similarity.
Performance
LFQA DPR-based retriever (vblagoje/dpr-question_encoder-single-lfqa-base and vblagoje/dpr-ctx_encoder-single-lfqa-base) had a score of 6.69 for R-precision and 14.5 for Recall@5 on KILT benchmark.
Usage
from transformers import DPRContextEncoder, DPRContextEncoderTokenizer
model = DPRQuestionEncoder.from_pretrained("vblagoje/dpr-question_encoder-single-lfqa-base").to(device)
tokenizer = AutoTokenizer.from_pretrained("vblagoje/dpr-question_encoder-single-lfqa-base")
input_ids = tokenizer("Why do airplanes leave contrails in the sky?", return_tensors="pt")["input_ids"]
embeddings = model(input_ids).pooler_output
Author
- Downloads last month
- 530