# wikipedia-dpr-dkrr-nq

Faiss FlatIP index of Wikipedia DPR encoded by the retriever model from [Distilling Knowledge from Reader to Retriever for Question Answering](https://arxiv.org/abs/2012.04584) trained on NQ.
This index was generated on 2022/02/17 on `orca` at commits:

+ Pyserini commit [`cc91b2`](https://github.com/castorini/pyserini/commit/cc91b22f549702068cea1283f91b31d28d127b2f) (2022/02/17)
+ [FiD](https://github.com/facebookresearch/FiD) commit [`25ed1f`](https://github.com/facebookresearch/FiD/commit/25ed1ff0fe0288b80fb5e9e5de8d6346b94b8d48) (2022/02/17)

with the following command to generate the embeddings (from FiD repo):

```bash
python generate_passage_embeddings.py \
  --model_path nq_retriever \
  --passages passages.tsv \
  --output_path wikipedia_embeddings_nq \
  --shard_id 0 \
  --num_shards 1 \
  --per_gpu_batch_size 500
```

and the following command to convert the embeddings to faiss IndexFlatIP form:

```bash
python convert_dkrr_embeddings_to_faiss.py \
  --embeddings wikipedia_embeddings_nq \
  --output faiss-flat.wikipedia.dkrr-dpr-nq-retriever
```