neoyipeng/modernfinbert-training-v2-long

ModernFinBERT Training Data v2 — Long Context

A long-context companion to modernfinbert-training-v2. Each row is a ~8,000-token passage from earnings call transcripts or SEC 10-K MD&A sections, with agent-labeled sentiment and entity annotations. Designed to leverage ModernBERT's 8,192-token context window for financial document understanding.

Dataset Summary

Property	Value
Total examples	36,025
Train / Val / Test	28,820 / 3,602 / 3,603
Labels	NEGATIVE, NEUTRAL, POSITIVE
Domains	Earnings call transcripts, SEC 10-K MD&A filings
Text length	2,000–32,000 chars (~500–8,000 tokens)
Unique entities	7,062
Entity coverage	85.9%
Unique companies	1,316
License	MIT

Label Distribution

Label	Count	Percentage
POSITIVE	19,890	55.2%
NEUTRAL	9,182	25.5%
NEGATIVE	6,953	19.3%

Source Composition

Source	Dataset	Rows	Domain	License
earnings_transcripts	glopardo/sp500-earnings-transcripts	21,500	Earnings calls	MIT
sec_10k_mda	jlohding/sp500-edgar-10k	14,525	SEC 10-K filings (Item 7 MD&A)	MIT

Label Distribution by Source

Source	NEGATIVE	NEUTRAL	POSITIVE
earnings_transcripts	973	3,545	16,982
sec_10k_mda	5,980	5,637	2,908

Earnings call transcripts skew positive (management tends to present optimistic narratives). MD&A sections are more balanced, with significant negative content (risk disclosures, challenges, regulatory concerns).

Schema

Column	Type	Description
text	string	Long financial passage (~500–8,000 tokens)
label	string	Overall passage sentiment: POSITIVE, NEGATIVE, or NEUTRAL
source	string	Source dataset: earnings_transcripts or sec_10k_mda
source_domain	string	Domain: earnings_calls or sec_filings
label_confidence	string	Always agent (all labels determined by Claude Code agents)
entity	string	Primary financial entity (company name, ticker, index, commodity, MARKET, or NONE)
entity_sentiment	string	Sentiment toward the specific entity

How This Dataset Was Produced

1. Source Data

• Earnings transcripts: 20,681 full S&P 500 earnings call transcripts from glopardo/sp500-earnings-transcripts, covering multiple years and quarters.
• SEC 10-K MD&A: Item 7 (Management's Discussion and Analysis) sections from 6,282 S&P 500 annual filings via jlohding/sp500-edgar-10k.

2. Chunking

Full transcripts (13K tokens) and MD&A sections (5K–50K tokens) were split into chunks of up to 32,000 characters (8,000 tokens) to fit within ModernBERT's 8,192-token context window. Chunks shorter than 2,000 characters (500 tokens) were discarded.

3. Deduplication

Exact text deduplication on normalized text (lowercased, whitespace collapsed) removed 4,821 duplicates from the 60,543 raw chunks, yielding 55,722 unique passages.

4. Balancing

Capped at ~43K rows to match the short-context v2 dataset. Balanced between sources: 21,500 earnings transcripts + 14,525 MD&A sections = 36,025 total.

5. Agent Annotation

All labels were determined by Claude Code agent workers (no pre-existing labels in source data). 361 batches of 100 rows each were processed by 8+ parallel agents. Each agent read the text passage and determined:

• label: overall passage sentiment
• entity: primary financial entity mentioned
• entity_sentiment: sentiment toward that entity

After annotation, entities were text-verified: every non-NONE/MARKET entity was confirmed to appear in its text via substring matching. 3,188 hallucinated entities were reset to NONE. Final entity coverage: 85.9%.

6. Split

80/10/10 stratified split by source. Zero cross-split text leakage verified.

Top Entities

Entity	Count
MARKET	4,301
General Electric Company	3,417
Ally Financial Inc.	1,267
Bank of America	1,047
General Motors Company	591
Goldman Sachs	493
Morgan Stanley	463
Bank of America Corp.	458
JPMorgan Chase & Co.	439
Goldman Sachs Group Inc.	346

Intended Use

• Fine-tuning ModernBERT (or other long-context encoders) for long-document financial sentiment classification
• Training alongside modernfinbert-training-v2 (short-context) for a model that handles both headlines and full documents
• Entity-aware financial sentiment research on long-form text

Limitations

• Agent-labeled: All labels were determined by AI agents, not human annotators. May contain systematic biases.
• Positive skew in earnings calls: Management presentations are inherently optimistic; the positive label dominance (55%) reflects this bias rather than balanced sentiment.
• Chunking artifacts: Some passages may start/end mid-sentence at chunk boundaries.
• Entity duplication: Some entities appear in both canonical and variant forms (e.g., "Bank of America" and "Bank of America Corp.") due to different agent workers.
• English only: All text is in English.
• S&P 500 bias: Both sources cover S&P 500 companies only; smaller companies are not represented.

Companion Dataset

This dataset is designed to be used alongside neoyipeng/modernfinbert-training-v2 (43K rows of short-context financial text: headlines, tweets, analyst reports). Combined training produces a model that handles both short and long financial text.

Citation

@dataset{glopardo2024,
  title={SP500 Earnings Transcripts},
  author={glopardo},
  year={2024},
  url={https://huggingface.co/datasets/glopardo/sp500-earnings-transcripts}
}

@dataset{jlohding2024,
  title={SP500 EDGAR 10-K},
  author={jlohding},
  year={2024},
  url={https://huggingface.co/datasets/jlohding/sp500-edgar-10k}
}