thelets/opendoor-knowledge-graph

Opendoor Financial Event Knowledge Graph — pipeline data

The gitignored data/ tree of the Prototyping-Financial-Knowlege-Graph repository, published so a second machine can regenerate the graph without re-fetching EDGAR.

This dataset is the data half only. The code, configuration, ontology and recorded provider answers live in the repository. Read Reproducing the graph before use — at the commit this snapshot was taken from, the repository alone cannot project this run, and the reason is recorded below rather than left to be discovered.

Anchor company: Opendoor Technologies Inc. (CIK 0001801169, Nasdaq: OPEN). Source: SEC EDGAR.

What is here

994.5 MiB across 2,755 files (1,042,852,388 bytes). Every path keeps its repository-relative name, so downloading into the repository root puts each file where the pipeline expects it.

Path	Size	Contents	Role
data/raw/sec/	746 MB	2,128 files — SEC source bytes, original filenames, per-filing _filing.json	Input to normalization
data/normalized/sec/	105 MB	588 files — per-document JSON + passages JSONL	Input to extraction
data/normalization_catalog/	41 MB	passages.jsonl, documents.jsonl, selection.jsonl, issues.jsonl	Input to the graph
data/extraction_runs/extract-v1-lexical-7f72d6172630/	33 MB	7 catalogs + manifest.json + lane_outputs.jsonl + report.md + run.complete	Input to the graph
data/graph_runs/graph-v1-eeba753149e9/	74 MB	nodes.jsonl, edges.jsonl, manifest.json	Graph output — for Neo4j load and byte-comparison
data/catalog/	1.9 MB	filings.jsonl, artifacts.jsonl	Derived acquisition index
data/normalization_reports/, data/normalization_runs/, data/reports/, data/runs/	1.4 MB	Per-run reports and issue files	Diagnostics
provenance/ontology_definitions/	88 KB	All 12 ontology YAML files as they were when the run was produced	See Reproducing the graph
provenance/CODE_STATE.txt	4 KB	Commit, branch, and the uncommitted-file list at snapshot time	Provenance
MANIFEST.sha256	250 KB	sha256 size path for all 2,755 files, sorted by path	Transfer verification

data/ is only two files away from being the graph's whole input: graph.core.inputs.load_run opens the extraction run directory and normalization_catalog/{passages,documents}.jsonl, and nothing else. Everything above those two rows exists so earlier stages can be replayed.

Deliberately excluded

Excluded	Size	Why
data/superseded_runs/extract-v1-lexical-2422c4252c07/	33 MB	Superseded extraction run
data/graph_runs/graph-v1-886059d862ce/	74 MB	Projected from that superseded run (projection 1.1.0)
.env	—	Local Neo4j password. .env.example is in the repository

data/ was scanned for credential patterns before upload; the only provider record in any manifest is a local model filename (Qwen3.5-9B-Q4_K_M.gguf) with provider_calls_permitted: 0.

Corpus and run facts

All values below are read from the shipped manifests, not restated from repository docs.

Layer	Measured
Acquisition	109 filings, 2,019 artifacts
Normalization	294 documents, 12,442 passages (1,935 table); corpus_id 294d-12442p-9992c4eb1888
Extraction	run extract-v1-lexical-7f72d6172630, created 2026-08-03T13:33:47Z
— candidates	11,848 (153 yielded a claim, 11,695 did not)
— claims / observations	2,714 / 2,704
— events / relationships	6 / 4
— issues / rejected claims	17,130 / 49
Graph	run graph-v1-eeba753149e9, projection 1.2.0, created 2026-08-03T13:34:46Z
— nodes	28,836 — Issue 17,127, Passage 8,776, Observation 2,704, Document 185, Metric 26, Entity 9, Event 6
— edges	35,600 — FOUND_IN 17,130, PART_OF 8,776, EVIDENCED_BY 2,710, HAS_OBSERVATION 2,704, OBSERVATION_OF_SUBJECT 2,704, CONCERNS_METRIC 1,523, DISTINCT_FROM 36, PARTICIPATES_IN 10, others 7

Ontology: real_estate_marketplace_v1, definition_hash 337e0e59534da0d2eb3235152af3d48f0510cdb391feed27956a4e80b0858af2. Extraction code commit: c59818e4a775103c94f3438306c051b66ef54387.

The extraction run replays 15 recorded provider answers rather than calling a model (mode: replay_only). The answers are committed in the repository at benchmarks/extraction/v1/answers/, so extraction is reproducible offline — no API key, no model download.

Getting the data

The repository is private, so a read token is required (https://huggingface.co/settings/tokens).

pip install huggingface_hub
hf auth login                     # paste a read token

cd /path/to/Prototyping-Financial-Knowlege-Graph
hf download thelets/opendoor-knowledge-graph --repo-type=dataset --local-dir .

--local-dir . writes data/... directly into the repository root, which is where config/*.yaml expects it. provenance/, README.md and MANIFEST.sha256 land beside it and are not read by any pipeline stage.

Verify the transfer

sha256sum -c MANIFEST.sha256

2,755 lines, all OK. The manifest is independent of the per-run manifests, which carry their own catalog_digests — checking both proves the run is internally consistent and that it arrived whole.

Reproducing the graph

The graph projection will refuse to run against a clean clone of commit c59818e. This is a measured fact, not a caution:

graph/core/manifest.py:171 compares the loaded ontology's definition_hash against the one the extraction run recorded and raises GraphInputError — "the vocabulary changed under a finalized run" — when they differ. They differ:

Ontology source	definition_hash
The tree that produced this run	337e0e59… (what manifest.json records)
Commit c59818e as committed	e8d4af70…

The run was produced from a working tree carrying uncommitted changes to ontology/versions/real_estate_marketplace_v1/definitions/{claims,constraints,events}.yaml (+80 lines) and to graph/core/models.py, where GRAPH_PROJECTION_VERSION is 1.2.0 uncommitted against 1.1.0 at c59818e. Since the graph run id derives from (projection_version, extraction_run_id, ontology_definition_hash) (graph/core/manifest.py:137), a clean clone would compute a different run id even if the hash gate were passed.

Two ways forward:

1. Preferred — publish the code. Commit and push branch impl/f0-factual-spine-contracts, then check it out on the second machine. The ontology YAML and the projection-version bump travel with it, both discrepancies disappear, and graph-v1-eeba753149e9 reprojects byte-for-byte.
2. Interim. provenance/ontology_definitions/ holds all 12 definition YAML files as they were when the run was produced — the hash covers the whole directory, so the full set ships rather than only the three that changed. Copying them over ontology/versions/real_estate_marketplace_v1/definitions/ restores definition_hash 337e0e59… and satisfies the gate. This does not restore the GRAPH_PROJECTION_VERSION bump, so the projection succeeds but mints a run id that is not graph-v1-eeba753149e9. Use it to read the shipped graph, not to claim reproduction.

Loading the shipped nodes.jsonl / edges.jsonl into Neo4j needs neither workaround — that output is already projected.

Provenance and licensing

Source documents are SEC EDGAR filings by Opendoor Technologies Inc., retrieved 2026-07-31. US government edicts and EDGAR submissions are not subject to copyright in the United States; the filings themselves are public disclosure. Raw bytes and original filenames are preserved exactly, and every artifact carries a hash, a source URL and a fetch timestamp.

The derived layers — normalized passages, extraction catalogs, the ontology, and the projected graph — are the work of this project. license: other is declared because the archive mixes public-source documents with project-authored derivations rather than because any part is restricted.

Extraction is imperfect by construction: this run holds 17,130 issues and 49 rejected claims, and rejected_claims.jsonl carries pre-validation model output that was refused. Nothing in rejected_claims.jsonl should be read as fact. None of this is investment advice.