---
license: other
---

# GraphSearch TAG datasets (processed)

Processed text-attributed-graph datasets used by the GraphSearch unified-agent project
(Amazon-Sports, Cora, RedditS). Per dataset:

- `<name>_corpus.json` — node records: id, split, text, 1-hop neighbors, label, degree, ppr_neighbors, semantic_neighbors
- `<name>_sbert_embeddings.pt` — all-mpnet-base-v2 node embeddings
- `<name>_ppr_top5.pt` — APPNP-approximated personalized PageRank (top-5)
- `<name>_test_ids.txt` — evaluation node ids (comma-separated)
- raw source files where applicable (csv / graph.pt)

Built with the pipeline in the GraphSearch repo: `data_utils/build_corpus.py` ->
`update_ppr_to_corpus.py` -> `encoder.py`.