Athithya117's picture
max: scale snapshot
62c44de verified
|
Raw History Blame Contribute Delete
1.83 kB
metadata
license: cc0-1.0
language:
  - en
tags:
  - legal
  - us-case-law
  - rag
  - graph-rag
  - citation-graph
  - adaptive-retrieval
size_categories:
  - 1K<n<10K
configs:
  - config_name: corpus_opinions
    data_files:
      - split: train
        path: corpus/opinions.parquet
  - config_name: graph_citations
    data_files:
      - split: train
        path: graph/citations_filtered.parquet
  - config_name: metadata_clusters
    data_files:
      - split: train
        path: metadata/clusters_filtered.parquet
  - config_name: eval_hojicha_corpus
    data_files:
      - split: train
        path: eval_isolated/hojicha/corpus.parquet
  - config_name: eval_hojicha_questions
    data_files:
      - split: train
        path: eval_isolated/hojicha/questions.parquet
  - config_name: eval_synthesized_graph_qa
    data_files:
      - split: train
        path: eval_isolated/synthesized_graph_qa/queries.parquet

adaptive-graphrag-datasets

US CourtListener GraphRAG dataset snapshot (scale=max).

See DATA_SOURCES.md for attribution and MANIFEST.json for SHA-256.

Scale presets

Scale Corpus rows Approx. vectors (@ ~5 chunks/doc)
min 5,000 ~25K
medium 100,000 ~500K
max 400,000 ~2M

Graph and metadata are filtered to the sampled corpus, so they scale with it.

Join keys

corpus.opinions.cl_opinion_id  <->  graph.citations.citing_opinion_id
corpus.opinions.cl_opinion_id  <->  graph.citations.cited_opinion_id
corpus.opinions.cluster_id     <->  metadata.clusters.id

metadata.clusters.docket_id    ->   dockets.id   (not shipped)
dockets.court_id               ->   courts.id    (court metadata is on corpus)

Quick start

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="<your-username>/adaptive-graphrag-datasets-max",
    repo_type="dataset",
    local_dir="data/raw",
)