INF-Retriever-v1

Model Overview

• INF-Retriever-v1 is an LLM-based dense retrieval model developed by INF TECH. It is built upon the gte-Qwen2-7B-instruct model and specifically fine-tuned to excel in retrieval tasks, particularly for Chinese and English data.

• As of December 23, 2024, INF-Retriever-v1 ranks No.1 on the Automated Heterogeneous Information Retrieval Benchmark of version 24.04(AIR-Bench_24.04), showcasing its cutting-edge performance in heterogeneous information retrieval tasks.

Key Features

• Optimized for Chinese and English retrieval: The model has been specifically fine-tuned with retrieval-focused datasets in both languages, significantly improving its accuracy and efficiency for a variety of retrieval scenarios.

• Top-tier performance: INF-Retriever-v1 has achieved outstanding results on the AIR-Bench leaderboard, making it a top choice for heterogeneous information retrieval tasks across various domains.

Usage

Sentence Transformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("infly/inf-retriever-v1", trust_remote_code=True)
# In case you want to reduce the maximum length:
model.max_seq_length = 8192

queries = [
    "how much protein should a female eat",
    "summit define",
]
documents = [
    "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
    "Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments.",
]

query_embeddings = model.encode(queries, prompt_name="query")
document_embeddings = model.encode(documents)

scores = (query_embeddings @ document_embeddings.T) * 100
print(scores.tolist())
# [[86.8702392578125, 67.82366180419922], [59.5101432800293, 82.33667755126953]]

Transformers

import torch
import torch.nn.functional as F

from torch import Tensor
from transformers import AutoTokenizer, AutoModel


def last_token_pool(last_hidden_states: Tensor,
                 attention_mask: Tensor) -> Tensor:
    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
    if left_padding:
        return last_hidden_states[:, -1]
    else:
        sequence_lengths = attention_mask.sum(dim=1) - 1
        batch_size = last_hidden_states.shape[0]
        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]


def get_detailed_instruct(task_description: str, query: str) -> str:
    return f'Instruct: {task_description}\nQuery: {query}'


# Each query must come with a one-sentence instruction that describes the task
task = 'Given a web search query, retrieve relevant passages that answer the query'
queries = [
    get_detailed_instruct(task, 'how much protein should a female eat'),
    get_detailed_instruct(task, 'summit define')
]
# No need to add instruction for retrieval documents
documents = [
    "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
    "Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments."
]
input_texts = queries + documents

tokenizer = AutoTokenizer.from_pretrained('infly/inf-retriever-v1', trust_remote_code=True)
model = AutoModel.from_pretrained('infly/inf-retriever-v1', trust_remote_code=True)

max_length = 8192

# Tokenize the input texts
batch_dict = tokenizer(input_texts, max_length=max_length, padding=True, truncation=True, return_tensors='pt')
outputs = model(**batch_dict)
embeddings = last_token_pool(outputs.last_hidden_state, batch_dict['attention_mask'])

# normalize embeddings
embeddings = F.normalize(embeddings, p=2, dim=1)
scores = (embeddings[:2] @ embeddings[2:].T) * 100
print(scores.tolist())
# [[86.87025451660156, 67.82366180419922], [59.510135650634766, 82.33667755126953]]

Evaluation

AIR-Bench

INF-Retriever-v1 has demonstrated superior retrieval capabilities across multiple domains and languages. The results from the Automated Heterogeneous Information Retrieval Benchmark of version 24.04(AIR-Bench_24.04) as of December 23, 2024, are as follows:

Model Name	Average	wiki_en	wiki_zh	web_en	web_zh	healthcare_en	healthcare_zh	law_en	arxiv_en	news_en	news_zh	finance_en	finance_zh	msmarco_en
BGE-M3	46.65	60.49	62.36	47.35	50.38	49.1	42.38	26.68	40.76	48.04	40.75	51.52	32.18	54.4
BGE-Multilingual-Gemma2	46.83	63.71	67.3	50.38	53.24	47.24	42.13	22.58	23.28	50.91	44.02	49.3	31.6	63.14
GTE-Qwen2-7B-instruct	48.38	63.46	66.44	51.2	51.98	54.2	38.82	22.31	40.27	54.07	43.03	58.2	26.63	58.39
INF-Retriever-v1	52.56	65.25	68.44	52.13	56.6	56.96	42.03	34.51	50.62	53.32	50.02	58.34	35.42	59.64