firqaaa/indo-dpr-ctx_encoder-multiset-base

indo-dpr-question_encoder-multiset-base

Indonesian Dense Passage Retrieval trained on translated SQuADv2.0 and Natural Question dataset in DPR format.

Evaluation

Class	Precision	Recall	F1-Score	Support
hard_negative	0.9961	0.9961	0.9961	384778
positive	0.8783	0.8783	0.8783	12414

Metric	Value
Loss	0.0220
Accuracy	0.9924
Macro Average	0.9372
Weighted Average	0.9924
Accuracy and F1	0.9353
Average Rank	0.2194

Note: This report is for evaluation on the dev set, after 27288 batches.

Usage

from transformers import DPRContextEncoder, DPRContextEncoderTokenizer

tokenizer = DPRContextEncoderTokenizer.from_pretrained('firqaaa/indo-dpr-ctx_encoder-multiset-base')
model = DPRContextEncoder.from_pretrained('firqaaa/indo-dpr-ctx_encoder-multiset-base')
input_ids = tokenizer("Siapa nama tokoh utama dalam serial SLAMDunk?", return_tensors='pt')["input_ids"]
embeddings = model(input_ids).pooler_output

You can use it using haystack as follows:

from haystack.nodes import DensePassageRetriever
from haystack.document_stores import InMemoryDocumentStore

retriever = DensePassageRetriever(document_store=InMemoryDocumentStore(),
                                  query_embedding_model="firqaaa/indo-dpr-ctx_encoder-multiset-base",
                                  passage_embedding_model="firqaaa/indo-dpr-ctx_encoder-multiset-base",
                                  max_seq_len_query=64,
                                  max_seq_len_passage=256,
                                  batch_size=16,
                                  use_gpu=True,
                                  embed_title=True,
                                  use_fast_tokenizers=True)