edge-dataset-artifacts / REPRODUCING.md
argsLiang's picture
Rename project to edge-dataset-artifacts and update documentation
68a2a08 verified
|
Raw History Blame Contribute Delete
7.65 kB

Edge Dataset Artifacts: reproduction and provenance

Release v1 decisions

Completeness is measured against the reference corpus, not directory size or date.

  • LihuaWorld graph: the historical minirag_v002 worktree's LiHuaWorld-MiniRAG-Graph-nothinking-qwen3.5-4B covers exactly all 425 week documents. April 18/19 candidates cover only 368/407. Older 442-document graphs contain 17 additional documents. The selected graph has 425 chunk vectors, 1,212 graph nodes and 507 edges. Its original build commit could not be recovered from the relocated worktree. Adjacent historical config is supporting evidence, not a guarantee of every historical runtime option.
  • ConcurrentQA graph: Conc-MiniRAG-Graph-nothinking-qwen3.5-4B-0430-keywordFIX covers all 2,743 documents. Other existing candidates cover 2,700, 2,708, or 2,740.
  • ConcurrentQA DiskANN: rebuilt from the 2,743 documents and existing 1,024-D embeddings in SQLite, without calling an embedding model. The old 18,551-document backup exactly matches the HF train split, shares only 1,288 texts with this test corpus, and is excluded.
  • Extended: Flat contains 991,397 vectors, with exactly 21 duplicate-text occurrences. Its 991,376 distinct texts exactly match SQLite and DiskANN. The original Flat vectors, IDs and duplicates are preserved. Scores between Flat and DiskANN are not claimed identical: duplicate multiplicity and ANN behavior differ.

All published FAISS indexes were loaded with their document stores and passed three self-query probes. GraphML, all vector matrix shapes/finite values, document coverage, chunk IDs and graph source references were checked. All DiskANN passage IDs and byte offsets, corpus coverage, required files, and PQ shapes were checked. Both DiskANN indexes also loaded successfully through the native backend. These checks do not establish end-to-end answer accuracy. Query-time embedding and generation still require model services.

Corpus reconstruction

LihuaWorld

The corpus archive installs data/LiHuaWorld/data/week*/*.txt and qa/query_set.{json,csv}. All 425 week files are byte-identical between the inspected old and new checkouts. bakery_deliver is excluded, matching the integrator. MiniRAG strips surrounding whitespace; Flat preserves the original text, so their text fingerprints differ even though document coverage matches. The supplied QA set has 637 entries. Existing placeholder evidence such as N/A and What does not resolve to TXT files; the integrator's warnings are preserved.

ConcurrentQA

  • HF dataset: stanfordnlp/concurrentqa.
  • Cached source revision: 2cc9db602b10a561026c905403a641422f066369.
  • Split: test, 1,600 questions, full range.
  • Traverse each question's sp in dataset order. For each nonempty title and sents, form title + "\n" + "\n".join(sents). Keep the first text for each title. Assign document IDs in insertion order. This gives 2,743 documents.
  • The original integrator downloads via HF and does not save a standalone corpus file. Its QA pairs refer to these document IDs; do not reorder documents when reconstructing them.

Example source download:

from datasets import load_dataset

dataset = load_dataset(
    "stanfordnlp/concurrentqa",
    revision="2cc9db602b10a561026c905403a641422f066369",
    split="test",
)

The online runner currently uses xrag's integrator and verifies the resulting evidence-text fingerprint. If upstream changes, use the source revision above; do not silently accept another corpus.

ConcurrentQA-Extended

This is a constructed dataset, not a single downloadable HF corpus:

  1. Start with the ConcurrentQA test documents above, preserving their IDs.
  2. Append parsed CMU Enron enron_mail_20150507 emails from https://www.cs.cmu.edu/~enron/enron_mail_20150507.tar.gz. The repository constructs text from Subject, Sender, Recipients, maildir path, a separator and the decoded body. Empty bodies and parse failures are skipped.
  3. Stream wikimedia/wikipedia, config 20231101.en, split train, shuffle seed 42, buffer size 10,000, and take 500,000 rows. Skip empty texts and repeated titles; rename titles colliding with base evidence using __ext. The locally cached Wikipedia revision is b04c8d1ceb2f5cd4588862100d08de323dccfbaa; the historical integrator did not pin a revision, so this is observed cache provenance, not proof of its exact input.
  4. Keep whole documents (identical_chunking) and drop texts longer than 20,000 characters. SQLite deduplicates by text hash. The stored release corpus has 991,376 distinct texts.

Maildir traversal used os.walk without sorting. Exact historical ID order cannot be guaranteed from source downloads alone, even with the same sampling seed. The released indexes retain their own mappings and texts. For byte-identical reproduction use the published artifacts; for source rebuilds compare corpus fingerprints and preserve the base QA document IDs.

Index and runtime configuration

  • Flat: FAISS IndexFlatIP, normalize_l2: true, 1,024 dimensions.
  • Flat/DiskANN embeddings: Qwen/Qwen3-Embedding-0.6B, truncation 2,048 tokens, local OpenAI-compatible provider. The release manifest records the complete xrag embedding configuration and the observed dependency versions.
  • MiniRAG: Qwen3-Embedding-0.6B, 1,024 dimensions; Qwen3.5-4B extraction. LihuaWorld predates later keyword-extraction fixes. Its larger coverage does not imply the same extraction behavior as the April 30 ConcurrentQA graph.
  • DiskANN: LEANN compact/recompute, inner product (mips), degree 32, build complexity 64, 8 threads, disk PQ 128 bytes, search/build memory budgets 8/16 GiB. The published leann.index is a prefix: the compact graph is stored in leann_disk_graph.index, not a file named leann.index.
  • Runtime DiskANN config matching directory hash 0ac1b1dfe7b9d7ca: num_nodes_to_cache: 1000, search_complexity: 16, beam_width: 4, prune_ratio: 0.5, pruning_strategy: global, batch_recompute: true, dedup_node_dis: true, recompute_neighbors: false. Use the audited configs/pipelines/eta/qwen/agentic_rag_leann_diskann.yml.
  • LEANN requires native DiskANN/partition support from xrag's pinned dependency checkout. A plain upstream diskannpy installation is not sufficient.
  • Compact DiskANN recomputes vectors at query time; the matching embedding model endpoint must be running. xrag refreshes the endpoint binding from your runtime configuration. Published metadata only contains localhost and the EMPTY placeholder.

The manifest's xrag/dependency commits identify the inspected code snapshot. They are not asserted to be the original build commits for historical artifacts. Model repository revisions and exact historical server settings were not recorded; the release therefore preserves artifacts rather than claiming bit-for-bit rebuilds.

Data terms

Index packages include document text in FAISS pickle stores, MiniRAG JSON, and LEANN passages. This release does not assign a new blanket license to those texts. Consult the original dataset cards and sources for applicable terms and attribution:

LihuaWorld is supplied from the xrag project's local corpus. Use the artifacts for the intended research reproduction and retain upstream attribution. Metadata, sample text, logs, and credentials beyond the required retrieval payload are not added to the release.