- build_trainset gains --radius (chars each side of the cited name) and --out;
merge-gap scales with radius. Default 600 unchanged.
- trainset_3x + train/val/test_3x.jsonl: same 10,519 triples and same trust split,
but ~3x more surrounding prose per triple (~47 -> ~132 tokens/triple, median
~3.7k tokens/sample). Keeps the 100% name-in-text guarantee.
- DATASET.md documents both context sizes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Track trainset/train/val/test.jsonl (focused excerpts only, no multi-MB text)
so the dataset can be used directly without the ~5h rebuild.
- DATASET.md: record schema, both serializations, load snippets (plain Python +
HF datasets), a text->triples fine-tuning sketch, eval notes, provenance.
- .gitignore: keep only the 100s-of-MB intermediates (samples_full) ignored.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- llm_extract.py: match mode now window-parallel with retrieval pre-filter,
claim dedup, retry, and enable_thinking=false (vLLM) -> ~36x faster per call;
n_failed_windows/ok flags so an interrupted run never records bogus 0s.
- build_rdf_dataset.py:
- gold now includes the share-class level (hasShareClass/ticker/className)
- grounding modes alias|llm|name|context|none (--grounding); llm reads the
role-check verdicts from match_all.jsonl
- label stage: per-triple extractable + per-sample FULL/PARTIAL/NONE
- trainset stage: combines GROUNDED triples with focused TEXT EXCERPTS cut
around the actual provider statement (evidence), not the multi-MB book
- split --src to split trainset.jsonl (trust-level, no leakage)
- helper scripts: watch_match.sh, resume_match.sh (crash/sleep-safe resume),
finalize_dataset.sh
- final dataset: 335/335 trusts, 85% text<->gold agreement, 334 samples,
10,689 grounded triples, train/val/test 264/35/35
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Ensures text<->gold agreement for the text->triple dataset:
- fetch: newest full prospectus book + later 497 supplements (time-aligned with
the current N-CEN gold; fixes stale 'largest book' picking 2-5yr-old filings)
- grounding: fast alias matcher (name present, variant-tolerant) AND an LLM
role-check (llm_extract.py match mode, via local Ollama or remote vLLM server)
that verifies the entity plays that ROLE -- catches right-name/wrong-role cases
a lexical matcher over-keeps. Validated with a strong model; ~93% of gold names
are present once granularity+currency are fixed.
- llm_extract.py: extract (baseline) + match (grounding) modes, sliding-window
with retrieval pre-filter, claim dedup, retry, ollama/vllm backends
- build_rdf_dataset.py: --grounding alias|llm|name|context|none, whole-trust
samples now filtered, MIN_PROSE stub guard
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Custodian names (esp. foreign sub-custodians) appear only in structured N-CEN,
never in the prospectus prose, so they are not a valid text->triple target.
Per-fund the custodian object name occurs in only 28% of segments, the weakest
of all relations. Default is now --custodian-scope none.
Every triple now carries a 'grounded' boolean (object name present in the
sample's input text); 80% of triples are grounded across the full build. This
lets training/eval restrict to text-extractable targets.
- build_rdf_dataset.py: annotate_grounding() + grounded flag in samples/stats
- gold rebuilt without custodian (15,739 -> 12,694 edges)
- dataset_description + README updated (custodian dropped, grounding documented)
Reported by thesis author: Citibank custodians in triples for 0001529390 never
appear in that prospectus text.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Builds a relationship-rich finance dataset for text-to-RDF-triple extraction
from SEC fund disclosures, the dataset for the thesis 'Magical RDF Triples and
how to synthetize them'.
- build_rdf_dataset.py: gold (N-CEN graphs), fetch (EDGAR prospectus prose,
all books per trust), samples (per-fund segmentation, marker + plain
serializations), split (trust-level 80/10/10, no leakage)
- score_baseline.py: no-model string-match baseline + strong-model scorer
- dataset_description.{tex,pdf}: scientific description of the dataset
- data/rdf_poc/gold_graphs.jsonl: structured gold knowledge graph (2025Q3)
- Large prose/sample files and raw SEC downloads are gitignored (reproducible)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
