bowang0911/edgar-8k-retrieval

EDGAR 8-K — Document-level Retrieval

Document-level (MTEB/BEIR-format) retrieval over SEC 8-K restructuring filings. Each query identifies the target filing by its event attributes — approximate date, sector, trigger/cause, and workforce-reduction size — with the company name, ticker, and unique product/person names removed, and asks for a fact the query does not state (e.g. identify the filing by its % workforce cut, ask for the $ restructuring charges). The model must distinguish among many near-identical restructuring filings by their specifics rather than by an entity-name lookup.

Example

"In May 2026, a business payments and financial-automation software company announced alongside its quarterly earnings that it would cut its workforce by up to 30% to improve organizational agility and drive profitability. What range of charges did the company estimate it would incur in connection with this restructuring?" → gold: BILL Holdings 8-K (2026-05-07).

Configs (MTEB format)

config	rows	fields
corpus	246	_id (accession), title (company + 8-K date), text (full filing)
queries	188	_id, text (event-based, name-free question)
default (qrels)	188	query-id, corpus-id, score

One gold document per query. Queries target restructuring charges or workforce-reduction facts. title in the corpus contains the company name for reference/analysis, but the queries never do.

Difficulty

Because queries omit the company name, this is a genuine content-matching task, not entity lookup. Baseline (all-MiniLM-L6-v2, CPU): recall@1 0.18 · recall@5 0.37 · recall@10 0.46 · nDCG@10 0.31 · MRR 0.27. (A name-based phrasing of the same queries scores ~0.97 recall@10 — the name is the entire shortcut, which this dataset removes.)

Construction & caveats

Built from 246 SEC 8-K restructuring filings (EDGAR full-text search, 2022–2026), cleaned to the Item narrative body. Event-based queries were generated by an LLM pass over each filing, instructed to describe the event without naming the company/ticker/product/person and to ask for a fact not used as the identifier; 9 queries that still leaked a distinctive company token were filtered out.

• LLM-paraphrased queries — natural but synthetic.
• Single gold per query — a few restructuring events are genuinely similar, so some misses reflect query ambiguity rather than model weakness.
• Coverage is 188 filings (one extraction batch failed; can be backfilled).

Source: public SEC EDGAR filings (U.S. government works / public domain).