Hugging Face Datasets - Financial Services

A shelf with real depth and no gatekeeper - which is simultaneously its charm and its hazard. Hugging Face Datasets - Financial Services is Datadory's governed feed over the machine-learning community's finance collections on the Hugging Face Hub: at the August 2026 research pass, about 19 repositories answer a 'financial services' query exactly, and the wider 'financial' searches clear 100. Row counts run from under a thousand to beyond a hundred million, and the named residents give the shelf its texture.

Three carry most of the weight. takala/financial_phrasebank holds 4,840 financial-news sentences annotated by 16 people with adequate background knowledge in financial markets, split into four agreement tiers so strictness becomes a filter rather than a compromise. zeroshot/twitter-financial-news-sentiment and its topic sibling supply text-label pairs harvested from finance Twitter. And defeatbeta/yahoo-finance-data bundles prices, earnings-call transcripts, SEC-filing text, dividend events and Treasury yields into a single repository banded above 100 million rows, with about 137,896 recorded downloads - the collection's workhorse.

Get a sample of this dataset and we return catalog rows plus payload records cut to the repositories your project actually names.

Two layers, deliberately kept apart. The catalog layer is one observation per repository and never changes shape; payload records arrive in whatever columns their uploader chose. From the August 2026 research pass:

# catalog observation - one row per repository
id            : defeatbeta/yahoo-finance-data
size_tag      : 100M<n<1B
downloads     : 137896          # cumulative at the August 2026 research pass

# payload observations - classification pairs from inside a repository
repo          : zeroshot/twitter-financial-news-sentiment
text          : $BYND - JPMorgan reels in expectations on Beyond Meat
label         : 0               # class index; readable mapping lives on the card

repo          : zeroshot/twitter-financial-news-sentiment
text          : $CCL $RCL - Nomura points to bookings weakness at Carnival and Royal Caribbean
label         : 0

# card excerpt - the shelf's most-cited sentiment benchmark
id            : takala/financial_phrasebank
excerpt       : collection of 4840 sentences ... annotated by 16 people
                with adequate background knowledge on financial markets
size_tag      : 1K<n<10K

Read the counter honestly: 137,896 downloads accumulates across versions, making it a popularity signal and nothing more - several heavily downloaded repositories turn out to be mirrors whose upstream sourcing cannot be traced. Note too what the label column does: 0 is an index, and the human-readable class meaning sits on each dataset card, which is exactly the decode work Datadory ships finished. (Catalog observations from the August 2026 research pass; payload records ship with your sample.)

Eight fields carry the catalog layer plus the commonest payload columns, defined below with examples taken from live records. id keys onto any model-asset registry you already maintain; task_categories and size_categories make the shelf queryable by intended job and declared heft; downloads and lastModified timestamp each repository's standing.

Everything else is per-repository and therefore folded rather than promised: the reuse declaration each uploader attaches to their card, the inner column sets behind the market-tabular and filing-excerpt repositories, and card configuration blocks get pinned with concrete examples when your sample names its targets.

• Geography: a global contributor base with English-dominant content - US and EU market material concentrates because that is where the harvested sources live - while country-level financial inclusion subsets push reach across Africa and Asia.
• Temporal: author-deposited snapshots rather than a maintained series. Observed last-modified stamps span early 2025 through August 2026, each repository carrying its own version history, and the matched set shifts as uploads land - any count on this page is a photograph, not a portrait.
• Granularity: one catalog observation per repository at the wrapper, descending to per-record grain inside them - sentence, tweet, filing excerpt or market observation - within independently versioned repositories.

Set against the wider Datadory catalog - where the average quality score across all 1,744 datasets is 7.81 - this slice scores 6/10: carried by breadth, named anchor corpora and fully-typed catalog columns, held back by unvetted community provenance and payload schemas that change from one repository to the next.

API, files, or your warehouse. Daily, weekly, or hourly.

Pick the channel your stack already speaks and set the cadence to match the decision being fed. Bulk files suit research benches loading whole corpora overnight; structured payloads suit products scoring headlines as they cross the tape; warehouse delivery suits teams joining these corpora to their own price and fundamentals tables without a second integration. The catalog layer, the common payload columns and the inner schemas of your chosen repositories travel unchanged across all three - cadence is a settings conversation, not a re-engineering project.

Five jobs it does better than any single official publisher.

Train and grade finance language models. Expert-labelled news sentences, tweet sentiment pairs and domain corpora sit one shelf apart, so tuning material and evaluation material arrive in one catalog filter instead of a procurement cycle.

Calibrate analyst-tone scoring. financial_phrasebank's four agreement tiers let you test a sentiment model against progressively stricter human consensus before trusting any vendor's number on a results-day headline.

Prototype document-understanding pipelines. OCR'd annual-report table collections and filing-derived corpora give extraction and retrieval work raw material that official publishers ship only as polished PDFs.

Screen the long tail before commissioning annotation. The discovery layer's modality, format, size-band and task filters turn 'what already exists' into a query - far cheaper to learn in August than after budgeting a labeling effort.

Stand up a market-data sandbox. yahoo-finance-data's single bundle - prices, transcripts, filings, dividends, Treasury yields - stocks a prototype environment with the joins already cohabiting one repository.

Data scientists and ML engineers get the shortest idea-to-experiment path in finance NLP: labelled text, typed catalog columns and versioned repositories. See data scientists use cases.

Investors and quants get the calibration set for tone: expert-agreement-tiered sentiment labels to hold their own scoring against, plus transcript and filing text for event studies. See investors quants use cases.

Developers and builders get stable repository identifiers and typed catalog rows to wrap into products, demos and retrieval features. See developers builders use cases.

Sales and growth teams selling into financial institutions get transcript and filing corpora that reveal what a prospect firm actually talks about - lending, payments, treasury - before the first call. See sales growth teams use cases.

Four things, stated plainly.

First, quality varies by uploader. Nobody curates this shelf; discipline ranges from institutional releases to weekend projects, and download counts will not tell you which is which. Sample preparation vets the repositories you name and flags the ones whose upstream sourcing cannot be traced.

Second, legacy packaging lingers. Some older repositories shipped with loading scripts that current tooling declines to execute; confirm a converted format exists before building anything on one, which is part of what a scoped sample settles.

Third, counts are photographs. Roughly 19 exact matches versus 100-plus wider ones reflects query width as much as inventory, and the matched set moves as uploads land - your sample states its own census date.

Fourth, the dictionary covers the wrapper plus the common payload columns. Inner schemas differ by design between a tweet pair and a hundred-million-row market bundle, and they get defined row by row once your sample fixes the repository list.

Notes that pair well with this page:

• NASDAQ Stock Screener API - the exchange's own listing universe, one row per security with sector and industry labels; the structured counterpart to community-built market bundles (profile).
• FINRA Data Catalog and API - regulator-grade broker-dealer transparency, where compliance work goes when community provenance is not enough (catalog).
• BIS Data Portal - Global Banking and Financial Statistics - central-bank statistics with institutional certainty behind every cell (portal).
• Diversified financial services data hub - the full pooled view of the industry, from ticker universes to bank regulators (hub).
• Best diversified-financial-services datasets - the ranked shortlist this slice sits within (ranking).