0xJupiter/SocialAttributionQA

Social Attribution QA Benchmark

The Social Attribution QA Benchmark is a derived benchmark for provenance-aware social attribution question answering over Fediverse data. It is designed to evaluate whether a system can identify who said a statement, what a person said, and whether attribution remains correct under entity, temporal, social, and collaborative constraints.

This release contains 1,200 four-option multiple-choice questions organized into eight task files. The benchmark is derived from the source dataset FediData and is released as a benchmark artifact rather than as a raw social-media dump.

This release is evaluation-oriented and is distributed as task files rather than as train/dev/test splits.

Dataset Summary

The benchmark is organized into two task families:

• WSW: Who Said What
• WDWS: What Did Who Say

Each JSON file contains a top-level dictionary with three fields:

• metadata: file-level provenance and construction metadata
• tasks: the benchmark instances for one task type
• statistics: counts and difficulty summaries for that task file

Data Files

File	Task	Questions
WSW_DIRECT.json	direct attribution	200
WSW_ENTITY.json	entity-constrained attribution	200
WSW_ASSOC.json	association reasoning	100
WSW_TEMPORAL.json	temporal attribution	100
WDWS_DIRECT.json	direct attribution	200
WDWS_ENTITY.json	entity-constrained attribution	200
WDWS_COLLAB.json	collaborative reasoning	100
WDWS_TEMPORAL.json	temporal attribution	100

Data Structure

Most instances contain the following fields:

• question_id: unique question identifier
• task_id: canonical task identifier
• question: question text
• options: four answer choices
• answer: gold option label such as A
• answer_text: gold answer in text form
• answer_path: supporting provenance information for the gold answer
• metadata: instance-level construction metadata
• difficulty: difficulty annotation and score

The collaborative file WDWS_COLLAB.json additionally includes correct_answer, while its difficulty annotation is not populated in the same way as the other task files.

Example

import json

with open("WSW_DIRECT.json", "r", encoding="utf-8") as f:
    data = json.load(f)

task_name = next(iter(data["tasks"]))
sample = data["tasks"][task_name][0]

print(task_name)
print(sample["question"])
print(sample["options"])
print(sample["answer"], sample["answer_text"])

Example task instance:

{
  "question_id": "WSW_T1_11c48887e878431b",
  "task_id": "WSW_T1_DIRECT",
  "question": "Who said: 'Smoking damages your lungs.'?",
  "options": {
    "A": "55ee6c1d@mastodon.social",
    "B": "bf0398ec@pouet.chapril.org",
    "C": "a25f92ab@mastodon.nl",
    "D": "ca4390cb@octodon.social"
  },
  "answer": "A",
  "answer_text": "55ee6c1d@mastodon.social"
}

Source Data

This benchmark is derived from the FediData Fediverse corpus:

• FediData: https://zenodo.org/records/15621244

This dataset repository does not redistribute the raw source-data dump. If you want to rebuild the benchmark from source, use the construction code in the project repository and place the downloaded FediData release under the expected build directory.

Related Resources

The full project repository includes:

• the released benchmark files
• the benchmark-construction pipeline
• baseline implementations
• the ATLAS method implementation

Project repository:

• https://github.com/JupiterXiaoxiaoYu/SocialAttributionQA

Intended Use

This release is intended for benchmark evaluation and method comparison. It is most suitable for:

• provenance-aware social attribution QA
• retrieval and reasoning over Fediverse-derived content
• comparison between graph-based, retrieval-based, and agentic QA methods

License

Apache License 2.0.