"तेल की कीमतें गिरकर 65 डॉलर प्रति बैरल हुईं" (Oil falls to $65) → Positive (India is an oil importer)
"भारतीय रिजर्व बैंक ने रेपो रेट में कटौती की" (RBI cuts repo rate) → Positive (stimulates growth)
"रुपया डॉलर के मुकाबले कमजोर हुआ" (Rupee weakens vs dollar) → Negative (hurts imports)

Part of the FLAME2 family.

Key Numbers


Language	Hindi
Market perspective	India (oil importer)
Training data	15,000 perspective-labeled headlines
Base model	IndicBERTv2-MLM-only (278M parameters)
Labels	Negative / Neutral / Positive
Accuracy	88.01%
F1 (macro)	88.01%

Quick Start

python

from transformers import pipeline

classifier = pipeline("text-classification", model="Kenpache/flame2-hindi")

# Oil prices fall — positive for India (importer)
classifier("तेल की कीमतें गिरकर 65 डॉलर प्रति बैरल हुईं")
# [{'label': 'positive', 'score': 0.94}]

# RBI cuts rate — positive
classifier("भारतीय रिजर्व बैंक ने रेपो रेट में 25 बीपीएस की कटौती की")
# [{'label': 'positive', 'score': 0.95}]

# Sensex drops — negative
classifier("सेंसेक्स 500 अंक गिरा, निवेशकों में बेचैनी")
# [{'label': 'negative', 'score': 0.93}]

# Company reports results — neutral
classifier("टाटा स्टील ने तिमाही नतीजे घोषित किए")
# [{'label': 'neutral', 'score': 0.88}]

No language prefix needed — this model is Hindi-only.

Results

Overall

Metric	Score
Accuracy	88.01%
F1 (macro)	88.01%

Per-Class Performance

Class	Precision	Recall	F1	Support
Negative	0.88	0.88	0.88	402
Neutral	0.88	0.84	0.86	709
Positive	0.88	0.93	0.90	640

Training Data


Total samples	15,000
Negative	3,543 (23.6%)
Neutral	5,902 (39.3%)
Positive	5,555 (37.0%)

Data sources include Indian financial news sites and economic news agencies. All headlines labeled from the Indian investor perspective — oil price drops are positive (India imports oil), rupee strengthening is positive, RBI rate cuts are positive.

Training Details

Parameter	Value
Base model	ai4bharat/IndicBERTv2-MLM-only (278M params)
Fine-tuning data	15,000 Hindi financial headlines
Loss function	Focal Loss (gamma=2.0)
Learning rate	2e-5 (→ 1e-5 SWA phase)
Label smoothing	0.1
Batch size	32
Max sequence length	128 tokens
Epochs	25
Precision	FP16 (mixed precision)
Train/Val/Test split	70% / 15% / 15%
SWA	Live averaging from epoch 12

Perspective Rules (India)

Event	Sentiment	Why
Oil prices fall	Positive	India is a major oil importer
Oil prices rise	Negative	Increases import costs
Rupee strengthens	Positive	Cheaper imports
Rupee weakens	Negative	Costlier imports
RBI rate cut	Positive	Stimulates economy
RBI rate hike	Negative	Tightens liquidity
Foreign CB actions	Neutral	Unless linked to Indian market

Batch Processing

python

from transformers import pipeline

classifier = pipeline("text-classification", model="Kenpache/flame2-hindi", device=0)

texts = [
    "सेंसेक्स ने 75000 का नया रिकॉर्ड बनाया",
    "महंगाई दर बढ़कर 6.5% पर पहुंची",
    "आरबीआई ने रेपो रेट में 25 बीपीएस की कटौती की",
    "रुपया डॉलर के मुकाबले 83.50 पर स्थिर",
    "रिलायंस का मुनाफा 15% बढ़ा",
]

results = classifier(texts, batch_size=32)
for text, result in zip(texts, results):
    print(f"{result['label']:>8} ({result['score']:.2f})  {text[:60]}")

Limitations

Optimized for news headlines (short text, 1-2 sentences)
Perspective reflects Indian economy — may not apply to Hindi speakers in other countries
Best on financial/economic news — general news or social media may underperform

Citation

bibtex

@misc{flame2_hindi_2026,
  title={FLAME2-Hindi: Financial Sentiment Analysis for Indian Markets},
  author={Kenpache},
  year={2026},
  url={https://huggingface.co/Kenpache/flame2-hindi}
}

License

Apache 2.0