andreiaalexa/adaptive-rag-scifact-50

SciFact — 50 Queries Sample

A fixed random sample of 50 queries from the SciFact test set, used as the evaluation set for the Adaptive RAG for Token-Efficiency course project.

Dataset Details

Field	Value
Source	BeIR/scifact
Split used	test
Sample size	50 queries
Random seed	0
Domain	Biomedical / Scientific fact-checking

Why this sample?

Running a full RAG pipeline with a local LLM (Mistral 7B via Ollama) over all 300 test queries is time-consuming. This fixed sample of 50 allows reproducible experiments across team members without rerunning the full set.

The seed is fixed at 0 — every team member gets the same 50 queries.

How to load

from datasets import load_dataset

ds = load_dataset("andreiaalexa/adaptive-rag-scifact-50")
queries = ds["queries"]

print(queries[0])
# {'_id': '...', 'text': '...'}

How it was created

import random
from datasets import load_dataset, DatasetDict

queries_ds = load_dataset("BeIR/scifact", "queries", split="queries")

random.seed(0)
indices = random.sample(range(len(queries_ds)), 50)
sampled = queries_ds.select(indices)

DatasetDict({"queries": sampled}).push_to_hub("andreiaalexa/adaptive-rag-scifact-50")

Project context

This dataset is part of the Adaptive RAG for Token-Efficiency project, developed for the Language Technology / Information Retrieval course. The project investigates whether adaptively selecting the number of retrieved documents per query can reduce token usage without hurting answer quality.

• Full project repository: (link to GitHub repo)
• Larger sample (150 queries): andreiaalexa/adaptive-rag-scifact-150

Citation

If you use SciFact, please cite the original dataset:

@inproceedings{wadden-etal-2020-fact,
    title = "Fact or Fiction: Verifying Scientific Claims",
    author = "Wadden, David and Lin, Shanchuan and Lo, Kyle and Wang, Lucy Lu and van Zuylen, Madeleine and Cohan, Arman and Hajishirzi, Hannaneh",
    booktitle = "Proceedings of EMNLP 2020",
    year = "2020"
}