Preprocessed artifacts for GAREN — Group-Aware Adaptive Retrieval for Evidence Navigation, covering all 12 BRIGHT categories. Everything here is derived from BRIGHT (Su et al., 2024).
With this dataset you can run GAREN directly, skipping the group-construction stage (which otherwise needs an 80B model and several hours of summarization).
| Path | Size | Description |
|---|---|---|
corpus/<task>.json | 2.2 GB | Chunked corpus, {docno, content} per document |
qa_data/<task>.json | 45 MB | Queries with gold_ids, excluded_ids, reasoning |
GRAPH/qwen3/graph/<task>/ | 209 MB | L0 doc–doc k-NN graph (k=16, Qwen3-Embedding-4B, exact cosine) |
GRAPH/qwen3/summary/<task>/ | 1.6 GB | L1 communities (Leiden) + LLM group summaries |
anchor/<task>/ | 164 MB | l0_l1_anchor.json (doc → L1 groups), l1_doc_ids.json (L1 group → docs) |
first_stage/bm25/ | 2.2 GB | BM25 first-stage retrieval caches (top-1000) |
Total: 6.7 GB.
Each GRAPH/qwen3/graph/<task>/ holds four files:
edges.u32.np uint32, (n_docs, k) row-major — neighbour indices
weights.f16.np float16, (n_docs * k,) — cosine similarities
docnos.npids npids.Lookup — row index ↔ docno
pt_meta.json {doc_count, k, ...}
| Task | Documents | Queries | L1 groups |
|---|---|---|---|
| biology | 57,359 | 103 | 5,183 |
| earth_science | 121,249 | 116 | 11,485 |
| economics | 50,220 | 103 | 4,561 |
| psychology | 52,835 | 101 | 4,789 |
| robotics | 61,961 | 101 | 4,861 |
| stackoverflow | 107,081 | 117 | 8,136 |
| sustainable_living | 60,792 | 108 | 5,540 |
| leetcode | 413,932 | 142 | 40,861 |
| pony | 7,894 | 112 | 621 |
| aops | 188,002 | 111 | 16,558 |
| theoremqa_questions | 188,002 | 194 | 16,558 |
| theoremqa_theorems | 23,839 | 76 | 2,196 |
| Total | 1,333,166 | 1,384 | 121,349 |
Qwen/Qwen3-Embedding-4B (vLLM pooling).faiss.IndexFlatIP after L2 normalization,
with self-edges removed.max_cluster_size=20), then one summary per community generated by
Qwen/Qwen3-Next-80B-A3B-Instruct-FP8.hf download Jun421/GAREN-data --repo-type=dataset --local-dir data/
Then, from the GAREN repository:
bash bash/run_ours.sh
The layout matches the paths the runner scripts expect, so no further moving is needed.
Preprocessed artifacts for GAREN — Group-Aware Adaptive Retrieval for Evidence Navigation, covering all 12 BRIGHT categories. Everything here is derived from BRIGHT (Su et al., 2024).
With this dataset you can run GAREN directly, skipping the group-construction stage (which otherwise needs an 80B model and several hours of summarization).
| Path | Size | Description |
|---|---|---|
corpus/<task>.json | 2.2 GB | Chunked corpus, {docno, content} per document |
qa_data/<task>.json | 45 MB | Queries with gold_ids, excluded_ids, reasoning |
GRAPH/qwen3/graph/<task>/ | 209 MB | L0 doc–doc k-NN graph (k=16, Qwen3-Embedding-4B, exact cosine) |
GRAPH/qwen3/summary/<task>/ | 1.6 GB | L1 communities (Leiden) + LLM group summaries |
anchor/<task>/ | 164 MB | l0_l1_anchor.json (doc → L1 groups), l1_doc_ids.json (L1 group → docs) |
first_stage/bm25/ | 2.2 GB | BM25 first-stage retrieval caches (top-1000) |
Total: 6.7 GB.
Each GRAPH/qwen3/graph/<task>/ holds four files:
edges.u32.np uint32, (n_docs, k) row-major — neighbour indices
weights.f16.np float16, (n_docs * k,) — cosine similarities
docnos.npids npids.Lookup — row index ↔ docno
pt_meta.json {doc_count, k, ...}
| Task | Documents | Queries | L1 groups |
|---|---|---|---|
| biology | 57,359 | 103 | 5,183 |
| earth_science | 121,249 | 116 | 11,485 |
| economics | 50,220 | 103 | 4,561 |
| psychology | 52,835 | 101 | 4,789 |
| robotics | 61,961 | 101 | 4,861 |
| stackoverflow | 107,081 | 117 | 8,136 |
| sustainable_living | 60,792 | 108 | 5,540 |
| leetcode | 413,932 | 142 | 40,861 |
| pony | 7,894 | 112 | 621 |
| aops | 188,002 | 111 | 16,558 |
| theoremqa_questions | 188,002 | 194 | 16,558 |
| theoremqa_theorems | 23,839 | 76 | 2,196 |
| Total | 1,333,166 | 1,384 | 121,349 |
Qwen/Qwen3-Embedding-4B (vLLM pooling).faiss.IndexFlatIP after L2 normalization,
with self-edges removed.max_cluster_size=20), then one summary per community generated by
Qwen/Qwen3-Next-80B-A3B-Instruct-FP8.hf download Jun421/GAREN-data --repo-type=dataset --local-dir data/
Then, from the GAREN repository:
bash bash/run_ours.sh
The layout matches the paths the runner scripts expect, so no further moving is needed.