neoyipeng/modernfinbert-training-v2

ModernFinBERT Training Data v2

A multi-source financial sentiment dataset for training financial sentiment models. Combines 5 source datasets spanning financial news, social media, analyst reports, and earnings call transcripts, with unified 3-class sentiment labels and agent-labeled entity-level sentiment annotations.

Dataset Summary

Property	Value
Total examples	43,247
Train / Val / Test	34,597 / 4,325 / 4,325
Labels	NEGATIVE, NEUTRAL, POSITIVE
Domains	Financial news, social media, analyst reports, earnings calls
Unique entities	4,209
Entity coverage	59.2%
License	ODC-By (Open Data Commons Attribution)

Label Distribution

Label	Count	Percentage
NEUTRAL	16,469	38.1%
POSITIVE	14,677	33.9%
NEGATIVE	12,101	28.0%

Source Composition

Source	Dataset	Rows	Domain	Original License
nosible	NOSIBLE/financial-sentiment	15,000	Financial news	ODC-By
timkoornstra_tweets	TimKoornstra/financial-tweets-sentiment	15,000	Social media	MIT
financemteb_finsent	FinanceMTEB/FinSent	9,929	Analyst reports	Not specified
subjectiveqa	gtfintechlab/SubjECTive-QA	2,621	Earnings calls	CC-BY 4.0
aiera_transcripts	Aiera/aiera-transcript-sentiment	697	Earnings calls	MIT

Large sources (NOSIBLE, TimKoornstra) were capped at 15,000 rows via label-stratified downsampling (5,000 per class). Small sources were kept at their natural size without upsampling. No maximum text length cap was applied -- ModernBERT supports up to 8,192 tokens.

Label Distribution by Source

Source	NEGATIVE	NEUTRAL	POSITIVE
nosible	5,000	5,000	5,000
timkoornstra_tweets	5,000	5,000	5,000
financemteb_finsent	1,825	4,545	3,559
subjectiveqa	210	1,498	913
aiera_transcripts	66	426	205

Schema

Column	Type	Description
text	string	Financial text to classify
label	string	Sentence-level sentiment: POSITIVE, NEGATIVE, or NEUTRAL
source	string	Source dataset identifier
source_domain	string	Domain category: financial_news, social_media, analyst_reports, or earnings_calls
label_confidence	string	Label provenance: human, human_aggregated, or llm_consensus
entity	string	Primary financial entity mentioned (company, ticker, index, commodity, sector, MARKET, or NONE)
entity_sentiment	string	Sentiment toward the specific entity (may differ from sentence-level label in multi-entity texts)

How This Dataset Was Produced

1. Collection & Label Alignment

Five source datasets were downloaded from HuggingFace and aligned to a unified 3-class schema (NEGATIVE / NEUTRAL / POSITIVE):

• NOSIBLE: Labels used as-is (positive/negative/neutral). No maximum text length cap applied.
• TimKoornstra: Mapped bullish -> POSITIVE, bearish -> NEGATIVE, neutral -> NEUTRAL
• FinanceMTEB FinSent: Labels used as-is (positive/negative/neutral)
• Aiera: Labels used as-is (positive/negative/neutral)
• SubjECTive-QA: Used the OPTIMISTIC dimension (0 -> NEGATIVE, 1 -> NEUTRAL, 2 -> POSITIVE) from the ANSWER column (earnings call responses)

Quality filters applied: minimum text length (10-20 chars depending on source), null removal.

2. Deduplication

Two-level deduplication removed 3,056 duplicates from the combined 151,339 rows:

1. Exact dedup: Normalized text (lowercased, whitespace collapsed), removed 95 exact duplicates
2. Near-duplicate dedup: MinHash with character 5-gram shingling (64 permutations, 16 LSH bands, Jaccard threshold > 0.8), removed 2,961 near-duplicates

Result: 148,283 unique rows.

3. Balancing

• Large sources (NOSIBLE, TimKoornstra) capped at 15,000 via label-stratified sampling (5,000 per class)
• Small sources (FinSent, SubjECTive-QA, Aiera) kept at natural size without upsampling
• Final dataset: 43,247 rows

4. Entity & Entity Sentiment Annotation

Entity extraction and entity-level sentiment were determined entirely by Claude Code agent workers in two passes. No external NLP packages (no spaCy, no NLTK) were used -- annotations are purely agent-determined.

Pass 1 -- Initial annotation: The balanced dataset (43,247 rows) was split into 217 batches of 200 rows. Multiple parallel agent workers read each batch, examined every text, and determined the entity and entity_sentiment using their financial domain knowledge.

Validation: After Pass 1, every non-NONE entity was verified to actually appear in its text using substring matching with canonical name variants. Entities that did not appear in the text (hallucinated by agents) were reset to NONE. This removed 17,247 incorrect annotations (67% of initial non-NONE entities).

Pass 2 -- NONE relabeling: The 34,397 rows marked NONE after validation were re-batched into 172 batches and sent to a second round of parallel agent workers for more careful annotation. This recovered 16,757 additional entity labels. After the same text-verification step, 4,281 hallucinated entities from Pass 2 were removed.

Final result: 25,607 rows (59.2%) have a verified entity, 17,640 rows (40.8%) remain NONE. 4,209 unique entities across the dataset. Every non-NONE entity is verified to appear in its text.

Entity breakdown:

• MARKET (general market commentary): 10,151 rows (23.5%)
• Named entities (companies, tickers, indices, commodities): 15,456 rows (35.7%)
• NONE (no identifiable entity): 17,640 rows (40.8%)

5. Train/Val/Test Split

80/10/10 stratified split on (source, label) pairs to preserve proportions across splits. Zero cross-split text leakage verified.

Intended Use

This dataset is designed for training and evaluating financial sentiment classification models, particularly:

• Fine-tuning encoder models (BERT, RoBERTa, ModernBERT, DeBERTa) for 3-class financial sentiment
• Benchmarking against held-out datasets: Financial PhraseBank, FiQA-SA, Twitter Financial News Sentiment
• Entity-aware financial sentiment analysis

Limitations

• Entity hallucination: Despite two-pass annotation with text verification, some edge cases may remain where entity names partially match unrelated text
• NONE coverage: 40.8% of rows have no identified entity, typically generic financial commentary or earnings call responses without explicit company names
• Label noise: The NOSIBLE source (34.7% of data) uses LLM-consensus labels, which may have systematic biases compared to human annotation
• Domain bias: Financial news and social media dominate (69.4%); earnings call coverage is limited (7.7%)
• English only: All text is in English
• Temporal coverage: Source datasets span different time periods; no temporal metadata is preserved in this version

Citation

If you use this dataset, please cite the original source datasets:

@dataset{nosible2025,
  title={NOSIBLE Financial Sentiment},
  author={NOSIBLE},
  year={2025},
  url={https://huggingface.co/datasets/NOSIBLE/financial-sentiment}
}

@dataset{timkoornstra2024,
  title={Financial Tweets Sentiment},
  author={Tim Koornstra},
  year={2024},
  url={https://huggingface.co/datasets/TimKoornstra/financial-tweets-sentiment}
}

@dataset{financemteb2024,
  title={FinSent},
  author={FinanceMTEB},
  year={2024},
  url={https://huggingface.co/datasets/FinanceMTEB/FinSent}
}

@inproceedings{subjectiveqa2024,
  title={SubjECTive-QA: Measuring Subjectivity in Earnings Call Transcripts},
  author={GTFinTechLab},
  year={2024},
  eprint={2410.20651},
  archivePrefix={arXiv}
}

@dataset{aiera2024,
  title={Aiera Transcript Sentiment},
  author={Aiera},
  year={2024},
  url={https://huggingface.co/datasets/Aiera/aiera-transcript-sentiment}
}