CiteTrace

761,495 citations from ten search-augmented LLMs answering 11,200 Stack Exchange questions across 28 sites, collected March 26–April 9, 2026.

Code

Load

from datasets import load_dataset

ds = load_dataset("anon-citetrace/CiteTrace", "analysis_master", split="train")

Replace analysis_master with any config below. Use revision= to pin a dataset version.

Files

Each config maps to data/<config>.parquet.

Join tables using query_id and url_id; cit_id identifies citations. Field metadata are in croissant.json. Score lookups and site definitions are in scoring_matrices/ and site_topology/.

Labels and validation

IPA scores use QI and SP; SS scores use SD and ST; ASF scores use the numeric label suffix. Classification validation contains 1,000 items and 3,000 human labels. Matrix validation contains 900 expert ratings: 30 IPA cells and 60 SS cells, each with ten scores. SS uses separate domain panels. Matrix tables provide scores, mean, median, and count.

Scope

The collection emphasizes technical and academic communities; model comparisons describe this collection window. Labels include unresolved conflicts and should not be treated as error-free ground truth. Source full text is not included. Classification consensus uses the first annotator when all three labels differ; pooled SS reliability combines different domain panels.

License

Stack Exchange queries: CC BY-SA 4.0, with attribution required. Research annotations and matrices: CC BY 4.0. Model responses and cited excerpts: CC BY 4.0 subject to provider terms. Code: MIT.