pipeline_tag	tags	license	language	metrics	datasets
feature-extraction	feature-extraction transformers	apache-2.0	id	accuracy f1 precision recall	squad_v2

indo-dpr-ctx_encoder-single-squad-base

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

Evaluation

Class	Precision	Recall	F1-Score	Support
hard_negative	0.9963	0.9963	0.9963	183090
positive	0.8849	0.8849	0.8849	5910

Metric	Value
Accuracy	0.9928
Macro Average	0.9406
Weighted Average	0.9928

Note: This report is for evaluation on the dev set, after 12000 batches.

Usage

from transformers import DPRContextEncoder, DPRContextEncoderTokenizer

tokenizer = DPRContextEncoderTokenizer.from_pretrained('firqaaa/indo-dpr-ctx_encoder-single-squad-base')
model = DPRContextEncoder.from_pretrained('firqaaa/indo-dpr-ctx_encoder-single-squad-base')
input_ids = tokenizer("Ibukota Indonesia terletak dimana?", 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-single-squad-base",
                                  passage_embedding_model="firqaaa/indo-dpr-ctx_encoder-single-squad-base",
                                  max_seq_len_query=64,
                                  max_seq_len_passage=256,
                                  batch_size=16,
                                  use_gpu=True,
                                  embed_title=True,
                                  use_fast_tokenizers=True)