spacemanidol/msmarco-v2.1-stella_en_1.5B_v5

NovaSearch stella_en_1.5B_v5 Embeddings for MSMARCO V2.1 for TREC-RAG

This dataset contains the embeddings for the MSMARCO-V2.1 dataset which is used as the corpora for TREC RAG All embeddings are created using Stella EN 1.5B V5 and are intended to serve as a simple baseline for dense retrieval-based methods. Note, that the embeddings are not normalized so you will need to normalize them before usage.

Retrieval Performance

Retrieval performance for the TREC DL21-23, MSMARCOV2-Dev and Raggy Queries can be found below with BM25 as a baseline. For both systems, retrieval is at the segment level and Doc Score = Max (passage score). Retrieval is done via a dot product and happens in BF16.

NDCG @ 10

Dataset	BM25	Stella-v5
Deep Learning 2021	0.5778	0.7285
Deep Learning 2022	0.3576	0.5568
Deep Learning 2023	0.3356	0.4875
msmarcov2-dev	N/A	0.3733
msmarcov2-dev2	N/A	0.3778
Raggy Queries	0.4227	0.5883
TREC RAG (eval)	N/A	0.6045

Recall @ 100

Dataset	BM25	Stella-v5
Deep Learning 2021	0.3811	0.42904
Deep Learning 2022	0.233	0.31683
Deep Learning 2023	0.3049	0.38541
msmarcov2-dev	0.6683	0.88111
msmarcov2-dev2	0.6771	0.871
Raggy Queries	0.2807	0.36869
TREC RAG (eval)	N/A	0.28572

Recall @ 1000

Dataset	BM25	Stella-v5
Deep Learning 2021	0.7115	0.74472
Deep Learning 2022	0.479	0.57709
Deep Learning 2023	0.5852	0.64342
msmarcov2-dev	0.8528	0.94629
msmarcov2-dev2	0.8577	0.94787
Raggy Queries	0.5745	0.66394
TREC RAG (eval)	N/A	0.68819

Loading the dataset

Loading the document embeddings

You can either load the dataset like this:

from datasets import load_dataset
docs = load_dataset("spacemanidol/msmarco-v2.1-stella_en_1.5B_v5", split="train")

Or you can also stream it without downloading it before:

from datasets import load_dataset
docs = load_dataset("spacemanidol/msmarco-v2.1-stella_en_1.5B_v5",  split="train", streaming=True)
for doc in docs:
    doc_id = j['docid']
    url = doc['url']
    text = doc['text']
    emb = doc['embedding']

Note, The full dataset corpus is ~ 620GB so it will take a while to download and may not fit on some devices/

Search

A full search example (on the first 1,000 paragraphs):

from datasets import load_dataset
import torch
from transformers import AutoModel, AutoTokenizer
import numpy as np


top_k = 100
docs_stream = load_dataset("spacemanidol/msmarco-v2.1-stella_en_1.5B_v5",split="train", streaming=True)

docs = []
doc_embeddings = []

for doc in docs_stream:
    docs.append(doc)
    doc_embeddings.append(doc['embedding'])
    if len(docs) >= top_k:
        break

doc_embeddings = np.asarray(doc_embeddings)


vector_dim = 1024
vector_linear_directory = f"2_Dense_{vector_dim}"
model = AutoModel.from_pretrained('NovaSearch/stella_en_1.5B_v5', trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained('NovaSearch/stella_en_1.5B_v5')
vector_linear = torch.nn.Linear(in_features=model.config.hidden_size, out_features=vector_dim)
vector_linear_dict = {
    k.replace("linear.", ""): v for k, v in
    torch.load(os.path.join(model_dir, f"{vector_linear_directory}/pytorch_model.bin")).items()
}
vector_linear.load_state_dict(vector_linear_dict)
model.eval() # ensure that model and vector linear are on the same device

query_prefix = 'Instruct: Given a web search query, retrieve relevant passages that answer the query.\nQuery: '
queries  = ['how do you clean smoke off walls']
queries_with_prefix = ["{}{}".format(query_prefix, i) for i in queries]
query_tokens = tokenizer(queries_with_prefix, padding=True, truncation=True, return_tensors='pt', max_length=512)

# Compute token embeddings
with torch.no_grad():
    attention_mask = *query_token["attention_mask"]
    last_hidden_state = model(***query_token)[0]
    last_hidden = last_hidden_state.masked_fill(~attention_mask[..., None].bool(), 0.0)
    query_vectors = last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
    query_vectors = normalize(vector_linear(query_vectors).cpu().numpy())


doc_embeddings = torch.nn.functional.normalize(doc_embeddings, p=2, dim=1)

# Compute dot score between query embedding and document embeddings
dot_scores = np.matmul(query_embeddings, doc_embeddings.transpose())[0]
top_k_hits = np.argpartition(dot_scores, -top_k)[-top_k:].tolist()

# Sort top_k_hits by dot score
top_k_hits.sort(key=lambda x: dot_scores[x], reverse=True)

# Print results
print("Query:", queries[0])
for doc_id in top_k_hits:
    print(docs[doc_id]['doc_id'])
    print(docs[doc_id]['text'])
    print(docs[doc_id]['url'], "\n")