dwzhu/LongEmbed

Introduction

This repo contains the LongEmbed benchmark proposed in the paper LongEmbed: Extending Embedding Models for Long Context Retrieval. Dawei Zhu, Liang Wang, Nan Yang, Yifan Song, Wenhao Wu, Furu Wei, Sujian Li, arxiv 2024.04. Github Repo for LongEmbed: https://github.com/dwzhu-pku/LongEmbed.

LongEmbed is designed to benchmark long context retrieval. It includes two synthetic tasks and four real-world tasks, featuring documents of varying lengths and dispersed target information. It has been integrated into MTEB for the convenience of evaluation.

How to use it?

Loading Data

LongEmbed contains six datasets: NarrativeQA, QMSum, 2WikiMultihopQA, SummScreenFD, Passkey, and Needle. Each dataset has three splits: corpus, queries, and qrels. The corpus.jsonl file contains the documents, the queries.jsonl file contains the queries, and the qrels.jsonl file describes the relevance. To spefic split of load each dataset, you may use:

from datasets import load_dataset

# dataset_name in ["narrativeqa", "summ_screen_fd", "qmsum", "2wikimqa", "passkey", "needle"]
# split_name in ["corpus", "queries", "qrels"]
data_list = load_dataset(path="dwzhu/LongEmbed", name="dataset_name", split="split_name")

Evaluation

The evaluation of LongEmbed can be easily conducted using MTEB (>=1.6.22). For the four real tasks, you can evaluate as follows:

from mteb import MTEB
retrieval_task_list = ["LEMBSummScreenFDRetrieval", "LEMBQMSumRetrieval","LEMBWikimQARetrieval","LEMBNarrativeQARetrieval"]
output_dict = {}
evaluation = MTEB(tasks=retrieval_task_list)
#TODO load the model before evaluation
results = evaluation.run(model,output_folder=args.output_dir, overwrite_results=True, batch_size=args.batch_size,verbosity=0)
for key, value in results.items():
    split = "test" if "test" in value else "validation"
    output_dict[key] = {"ndcg@1": value[split]["ndcg_at_1"], "ndcg@10": value[split]["ndcg_at_10"]}
print(output_dict)

For the two synthetic tasks, since we examine a broad context range of {256, 512, 1024, 2048, 4096, 8192, 16384, 32768} tokens, an additional parameter of context_length is required. You may evaluate as follows:

from mteb import MTEB
needle_passkey_task_list = ["LEMBNeedleRetrieval", "LEMBPasskeyRetrieval"]
output_dict = {}
context_length_list = [256, 512, 1024, 2048, 4096, 8192, 16384, 32768]
evaluation = MTEB(tasks=needle_passkey_task_list)
#TODO load the model before evaluation
results = evaluation.run(model, output_folder=args.output_dir, overwrite_results=True,batch_size=args.batch_size,verbosity=0)
for key, value in results.items():
    needle_passkey_score_list = []
    for ctx_len in context_length_list:
        needle_passkey_score_list.append([ctx_len, value[f"test_{ctx_len}"]["ndcg_at_1"]])
    needle_passkey_score_list.append(["avg", sum([x[1] for x in needle_passkey_score_list])/len(context_length_list)])
    output_dict[key] = {item[0]: item[1] for item in needle_passkey_score_list}
print(output_dict)

Task Description

LongEmbed includes 4 real-world retrieval tasks curated from long-form QA and summarization. Note that for QA and summarization datasets, we use the questions and summaries as queries, respectively.

• NarrativeQA: A QA dataset comprising long stories averaging 50,474 words and corresponding questions about specific content such as characters, events. We adopt the test set of the original dataset.
• 2WikiMultihopQA: A multi-hop QA dataset featuring questions with up to 5 hops, synthesized through manually designed templates to prevent shortcut solutions. We use the test split of the length-uniformly sampled version from LongBench.
• QMSum: A query-based meeting summarization dataset that requires selecting and summarizing relevant segments of meetings in response to queries. We use the version processed by SCROLLS. Since its test set does not include ground truth summarizations, and its validation set only have 60 documents, which is too small for document retrieval, we include the train set in addition to the validation set.
• SummScreenFD: A screenplay summarization dataset comprising pairs of TV series transcripts and human-written summaries. Similar to QMSum, its plot details are scattered throughout the transcript and must be integrated to form succinct descriptions in the summary. We use validation set of the version processed by SCROLLS.

We also include two synthetic tasks, namely needle and passkey retrieval. The former is tailored from the Needle-in-a-Haystack Retrieval for LLMs. The later is adopted from Personalized Passkey Retrieval, with slight change for the efficiency of evaluation. The advantage of synthetic data is that we can flexibly control context length and distribution of target information. For both tasks, we evaluate a broad context range of {256, 512, 1024, 2048, 4096, 8192, 16384, 32768} tokens. For each context length, we include 50 test samples, each comprising 1 query and 100 candidate documents.

Task Statistics

Dataset	Domain	# Queries	# Docs	Avg. Query Words	Avg. Doc Words
NarrativeQA	Literature, File	10,449	355	9	50,474
QMSum	Meeting	1,527	197	71	10,058
2WikimQA	Wikipedia	300	300	12	6,132
SummScreenFD	ScreenWriting	336	336	102	5,582
Passkey	Synthetic	400	800	11	-
Needle	Synthetic	400	800	7	-

Citation

If you find our paper helpful, please consider cite as follows:

@article{zhu2024longembed,
  title={LongEmbed: Extending Embedding Models for Long Context Retrieval},
  author={Zhu, Dawei and Wang, Liang and Yang, Nan and Song, Yifan and Wu, Wenhao and Wei, Furu and Li, Sujian},
  journal={arXiv preprint arXiv:2404.12096},
  year={2024}
}