Datasets:

ethicalabs
/

Research-Intent-Collab

license: cc-by-4.0
task_categories:
  - text-classification
language:
  - en
  - multilingual
tags:
  - research-intent
  - golden-dataset
  - openaire
pretty_name: Research Intent — Collaborative Golden Dataset

Research Intent — Collaborative Golden Dataset

▶️ Watch the Video

A curated, consensus-labelled dataset for research paper intent classification, built from the LLM-as-Judge annotations of the Echo-DSRN collaborative platform (OpenAIRE AI Hackathon 2026).

OpenAIRE AI Hackathon 2026 🇪🇺

This dataset is part of ethicalabs.ai's entry in the OpenAIRE AI Hackathon 2026, co-organised by OpenAIRE and Alien Intelligence — a 12-week open science build challenge.

🔴 Live Demo → openaire-2026.ethicalabs.ai

The live application classifies research papers in real time using this model, with publication metadata streamed from the OpenAIRE Graph API under CC BY 4.0.

Labels

Methodology — introduces a new algorithm, model architecture, or method
Dataset — primarily contributes a new dataset, benchmark, or annotation scheme
Review — survey, systematic review, or position paper
Applied — applies existing methods to a domain problem
Theoretical — mathematical or formal analysis
Unclassifiable — garbled, non-English, or non-academic content

Golden labelling

The golden label per paper is the majority-vote consensus across all LLM judge models (plus human curator annotations where available). Papers flagged as low-quality by more than one judge are labelled Unclassifiable instead of being discarded — turning a quality filter into a detection signal for non-academic content.

Splits

The dataset ships with stratified train / val / test splits (80/10/10). Stratification preserves the label distribution across splits, with explicit safeguarding so the rare Dataset class (≈2% of records) appears in every split.

Data structure

Consensus label per paper — no ChatML messages and no judge rationale, so consumers build their own training format from title + description + label.

{
  "doi": "10.48550/arXiv.2106.01345",
  "title": "...",
  "description": "abstract text",
  "label": "Dataset",
  "flag_count": 0,
  "model_votes": {
    "Gemma-4-E4B-it-GGUF": "Dataset",
    "Qwen3.5-4B-GGUF": "Methodology"
  }
}

label — the golden consensus label
flag_count — number of judges that flagged quality issues
model_votes — per-model predictions used for the consensus

For ChatML conversation format with inline reasoning, see the companion dataset ethicalabs/Research-Intent-Judge.

Files

golden_train.jsonl — 7,999 records
golden_val.jsonl — 1,001 records
golden_test.jsonl — 1,001 records
golden_stats.json — label distribution, split sizes, flag counts

Provenance

Built from ethicalabs/Research-Intent-Judge via scripts/export_golden_dataset.py in the OpenAIRE-AI-Research-Evaluator repository. Paper metadata originates from arXiv and the OpenAIRE Graph (CC BY 4.0).

License

CC-BY 4.0.

Citation

If you reuse this work, please cite:

@software{echo_dsrn_research_intent_clf,
  author = {Massimo Roberto Scamarcia},
  title = {OpenAIRE Graph Evaluator: Research paper classifier and LLM-as-Judge pipeline},
  year = {2026},
  url = {https://huggingface.co/collections/ethicalabs/openaire-ai-hackathon-2026}
}