techkiyan/indian-financial-news-ner-gliner-v1

Indian Financial News NER GLiNER v1

indian-financial-news-ner-gliner-v1 is a GLiNER-based named entity recognition model fine-tuned for Indian financial news, market updates, company news, and stock-market-related text.

The model is designed as a lightweight candidate extraction layer for financial intelligence pipelines. It can be used to extract companies, indices, organizations, amounts, percentages, dates, sectors, financial terms, and event-related phrases from short financial-news text.

This is a v1 release. It should be treated as a domain-specialized candidate extractor, not as a final truth engine.

Model Details

• Base model: EmergentMethods/gliner_medium_news-v2.1
• Model family: GLiNER
• Task: Named Entity Recognition
• Domain: Indian financial news and market news
• Training type: supervised fine-tuning
• Hardware used: NVIDIA RTX 2060 12GB
• Training examples: 73,864
• Validation examples: 9,055
• Training steps: 18,500
• Training runtime: about 44 minutes

Labels

The model supports 10 labels:

• company
• index
• organization
• person
• amount
• percentage
• date
• financial_term
• event_signal
• sector

Label Definitions

company

Company mentions, including listed companies, IPO companies, and other business entities in financial-news context.

Examples:

• Reliance Industries
• HDFC Bank
• Infosys
• NSDL

index

Market indices and benchmarks.

Examples:

• Nifty 50
• Sensex
• Bank Nifty
• Nifty IT

organization

Regulators, exchanges, rating agencies, government bodies, courts, brokerages, and other financial organizations.

Examples:

• SEBI
• RBI
• NSE
• BSE
• CRISIL
• ICRA

person

Named individuals such as executives, promoters, ministers, regulators, analysts, and officials.

Examples:

• Shaktikanta Das
• Nirmala Sitharaman
• Mukesh Ambani

amount

Monetary values.

Examples:

• ₹1,800 crore
• Rs 500 crore
• $35 million

percentage

Percentages, basis points, and rate movement expressions.

Examples:

• 5%
• 25 bps
• 4.77 per cent

date

Dates, quarters, financial years, and time-period references.

Examples:

• Q1FY26
• FY25
• June 2026

financial_term

Financial metrics, instruments, corporate-action terms, and market-finance concepts.

Examples:

• revenue
• EBITDA
• NPA
• order book
• dividend
• bond
• buyback
• QIP

event_signal

Market or business event trigger phrases.

Examples:

• wins order
• receives approval
• raises funds
• reports profit growth
• board approves buyback
• resigns

sector

Industry or market sectors.

Examples:

• banking
• IT
• pharma
• NBFC
• auto
• energy

Usage

from gliner import GLiNER

model = GLiNER.from_pretrained("techkiyan/indian-financial-news-ner-gliner-v1")

labels = [
    "company",
    "index",
    "organization",
    "person",
    "amount",
    "percentage",
    "date",
    "financial_term",
    "event_signal",
    "sector",
]

text = "HDFC Bank shares rose 3% after the company reported strong Q1 profit growth."

entities = model.predict_entities(text, labels, threshold=0.5)

for entity in entities:
    print(entity)

Recommended Use

This model is best used as the first layer in a financial information extraction pipeline:

1. Extract entity candidates.
2. Resolve company names to NSE/BSE symbols.
3. Normalize amounts, percentages, and dates using deterministic rules.
4. Link organizations and sectors to internal dictionaries.
5. Use a separate event classifier for event type and market-intelligence workflows.

Recommended Labels for Production Pipelines

For stricter production use, start with:

• company
• index
• organization
• amount
• percentage
• financial_term
• sector

The following labels should be treated as experimental or review-only in v1:

• event_signal
• person
• date

Evaluation

Internal corrected merged-label test benchmark:

Metric	Value
Precision	0.9225
Recall	0.8850
F1	0.9033
Threshold	0.95

Important: this is an internal benchmark on corrected merged-label test data. It should not be interpreted as a public SOTA claim.

Dataset

The model was trained on a corrected silver + seed-gold Indian financial-news corpus.

Summary:

• Total exported rows: 92,485
• Annotation-corrected rows: 43,584
• Reviewed rows: 100
• Seed-gold rows: 95
• Training rows: 73,864
• Validation rows: 9,055

The training data is not included in this model release. Some source text may come from third-party financial-news sources, so users should verify licensing before redistributing datasets.

Limitations

• This is a v1 model.
• It is intended for candidate extraction, not final financial reasoning.
• Entity linking to NSE/BSE symbols is not included.
• event_signal is experimental and may be better handled by a separate event classification model.
• The model does not provide investment advice, trading signals, buy/sell calls, or price predictions.

Disclaimer

This model is for financial text understanding and market-intelligence research. It does not provide investment advice, buy/sell recommendations, target prices, or trading decisions.