Hugging Face's logo

Datasets:
barissozudogru
/
federated-search-traces

Tasks:
Text Retrieval
Modalities:
Tabular
Text
Formats:
json
Languages:
English
Size:
< 1K
Tags:
academic-search
federated-search
arxiv
pubmed
semantic-scholar
openalex
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
federated-search-traces / README.md
barissozudogru's picture
barissozudogru
Polish README: drop placeholder copy, add YAML cross-links to related artifacts
38aab8f verified
preview code
|
Raw
History Blame Contribute Delete
4.96 kB
metadata 
license: cc-by-4.0
language:
  - en
size_categories:
  - n<1K
task_categories:
  - text-retrieval
tags:
  - academic-search
  - federated-search
  - arxiv
  - pubmed
  - semantic-scholar
  - openalex
  - retrieval
  - benchmark
  - deduplication
  - research-papers-mcp
configs:
  - config_name: default
    data_files:
      - split: train
        path: traces.json

Federated Academic Search Traces (research-papers-mcp benchmark)

What this is

30 curated academic search queries run through the research-papers-mcp federated search pipeline. For each query, the dataset captures:

  • The query text + domain label + expected-strong-sources annotation
  • Per-source raw result counts (before dedup)
  • Total deduplicated count + number of duplicates removed
  • Top 10 deduplicated paper records (title, year, source, DOI, URL)
  • Per-source wall-clock timing

The queries span 30 domains across ML, biomedicine, physics, philosophy, economics, computer science, and astronomy — chosen to exercise each source's coverage strengths.

Sources active during this trace

This snapshot was captured against the research-papers-mcp package as installed from main branch at the time of run. The registry contained three live sources:

Source Status during trace
arXiv Active (rate-limited on bulk queries — see per_source_timing_sec for evidence)
PubMed Active, 30/30 queries returned results
Semantic Scholar Active but throttled (HTTP 429 on 23/30 queries)
OpenAlex Not registered on the main branch at trace time. A PaperSource-protocol package abstraction exists on a feature branch but had not been merged. Future traces will include OpenAlex once that branch lands.

This is itself a useful signal: the trace honestly records which sources a pip install -e . from main got at trace time. Users running newer versions of the package will see different (likely fuller) coverage.

Schema

Field Type Description
id string Query identifier (q01 to q30)
query string The search query text
domain string Coarse-grained topic domain
expected_strong_sources list[string] Sources we expect to return strong matches
per_source_raw_count dict[str, int] Records returned by each source before dedup
raw_total int Sum across sources
deduped_count int After DOI-then-(source, source_id) deduplication
duplicates_removed int raw_total - deduped_count
per_source_timing_sec dict[str, float] Wall-clock seconds per source
top_papers list[object] First 10 deduplicated results (title, year, source, doi, url)

Load it 

from datasets import load_dataset
ds = load_dataset("barissozudogru/federated-search-traces")
print(ds["train"][0])

Or directly:

import json
import requests
data = requests.get("https://huggingface.co/datasets/barissozudogru/federated-search-traces/resolve/main/traces.json").json()
for r in data[:5]:
    print(r["query"], "->", r["deduped_count"], "results")

Generating method

Queries were run through the research-papers-mcp pipeline at https://github.com/barissozudogru/deep-research-digest. Each source is called concurrently with max_results=15. Deduplication is by DOI first (lowercased, stripped), then by (source, source_id). Trace is captured at one point in time — re-running may give different counts as the live indices update.

The full reproduction script is run_benchmark.py in this dataset's repo.

Use cases

  • Regression-testing federated search dedup logic
  • Comparing source coverage across domains
  • Benchmarking new sources (e.g., does adding CORE or bioRxiv change dedup ratios meaningfully?)
  • Sanity-checking ranking algorithms on stable inputs

Caveats

  • Single snapshot, not a live mirror. External indices update; re-running may produce different counts.
  • Sources may rate-limit. If per_source_raw_count[name] == 0 for a query, that source either truly had no matches OR was throttled at trace time. Compare with per_source_timing_sec for hints.
  • Query selection is editorial, not random. We picked queries to exercise the four sources' coverage strengths. Inferences about typical user behavior should treat this as a stress-test corpus, not a representative one.
  • No ground-truth relevance labels. This is a trace dataset (what came back), not an annotated benchmark (what should have come back). For relevance evaluation, pair with human-annotated relevance judgments.

Related