README.md

---
license: apache-2.0
pretty_name: CVEfixes Security IR GraphRAG
configs:
- config_name: bm25_documents
  data_files:
  - split: train
    path: data/bm25/documents/*.parquet
- config_name: bm25_keyword_index
  data_files:
  - split: train
    path: indexes/bm25_keyword_shards.parquet
- config_name: bm25_postings
  data_files:
  - split: train
    path: data/bm25/postings/*.parquet
- config_name: corpus
  data_files:
  - split: train
    path: data/corpus/*.parquet
- config_name: corpus_chunk_index
  data_files:
  - split: train
    path: indexes/corpus_chunks.parquet
- config_name: graph_edges
  data_files:
  - split: train
    path: data/graph/edges/*.parquet
- config_name: graph_incoming_adjacency
  data_files:
  - split: train
    path: data/graph/adjacency/incoming/*.parquet
- config_name: graph_incoming_adjacency_index
  data_files:
  - split: train
    path: indexes/graph_incoming_adjacency.parquet
- config_name: graph_nodes
  data_files:
  - split: train
    path: data/graph/nodes/*.parquet
- config_name: graph_outgoing_adjacency
  data_files:
  - split: train
    path: data/graph/adjacency/outgoing/*.parquet
- config_name: graph_outgoing_adjacency_index
  data_files:
  - split: train
    path: indexes/graph_outgoing_adjacency.parquet
- config_name: original_data
  data_files:
  - split: train
    path: data/original/*.parquet
- config_name: original_row_index
  data_files:
  - split: train
    path: indexes/original_rows.parquet
- config_name: vector_meta_index
  data_files:
  - split: train
    path: indexes/vector_chunks.parquet
- config_name: vectors
  data_files:
  - split: train
    path: data/vectors/*.parquet
source_datasets:
- hitoshura25/cvefixes
---

# CVEfixes Security IR GraphRAG

This is a content-addressed, remotely routable Security IR release of the exact pinned CVEfixes snapshot. It packages the byte-identical original Parquet data together with a searchable corpus, BM25 postings, CUDA-generated vectors, typed graph nodes and edges, bounded adjacency indexes, and a source-CID-to-original-row lookup.

All entries are inert, non-authoritative evidence. Candidate and formal-logic rows cannot grant execution authority; exact forbidden action and scope constraints remain explicitly unresolved until reviewed by the Security IR policy workflow.

## Immutable build bindings

- Source: `hitoshura25/cvefixes@d4f5c4ea65329d9ccbb8a3b3149e5d06eda5edb2`
- Derived Security IR root: `bafkreia74ozdbgzwnirt7mixgfwwffiku24vbtwmivzyvvyghlgqhvkwk4`
- Graph root: `bafkreielsquxgqxh6qzb3444bqjlicl34fxqtkyjebwa5h3vqhtaygynee`
- Retrieval root: `bafkreigcqhkshkplcpc6k5ppufzqwh4x2xte544bqu4wbuvhluvth25r6e`
- Embedding model: `sentence-transformers/all-MiniLM-L6-v2@1110a243fdf4706b3f48f1d95db1a4f5529b4d41`
- CUDA device: `NVIDIA GB10`
- CUDA runtime: `13.0`
- Embedding dimension: `384`

## Coverage

- Source rows represented: `12987`
- Original rows packaged: `12987`
- Byte-identical original shards: `3`
- Admitted rows: `12714`
- Rejection tombstones: `273`
- Graph nodes: `85169`
- Graph edges: `167364`

## Remote index contract

Each physical Parquet shard is bound by SHA-256, raw CIDv1, byte size, row count, key range, and path. The Dataset Viewer exposes the declared data/meta configurations used by Publicus/skillcenter-ir.
The `original_row_index` config maps every Security IR source CID to a global row index, copied shard CID/path, and shard-local offset.

## Safety and limitations

- `original_data` is a byte-preserving mirror of the pinned upstream dataset and includes unfiltered diffs, vulnerable/fixed code, secret-like strings, personal data, unsafe paths, and rows excluded from the derived public Security IR. Treat every value as untrusted inert evidence; never execute it or use it as a credential.
- The upstream dataset card declares Apache-2.0. Embedded code comes from many repositories and may retain additional record-level license obligations; consumers must review those sources.
- Only the derived corpus/graph configs are policy-filtered and bounded. Original bodies are not injected directly into agent context.
- Graph, lexical, and vector similarity are retrieval evidence only.
- Classification-only candidates do not infer forbidden behavior.
- Consumers must pin a 40-character Hub commit before querying.