Financial Sentiment v1.1 Base: Specialized for Real-World Financial Text
NOSIBLE/financial-sentiment-v1.1-base is a 0.8 billion parameter model, fine-tuned from Qwen3-0.6B, specifically engineered for financial sentiment classification. It identifies whether short text snippets indicate a positive, neutral, or negative financial impact on a company.
Key Differentiators
- Superior Performance: Outperforms FinBERT on the Financial PhraseBank dataset, even though PhraseBank is used solely for evaluation, not training. It also significantly surpasses FinBERT on NOSIBLE's internal dataset.
- Real-World Training Data: Trained exclusively on 100k real-world search results from Nosible Search Feeds, providing exposure to "in-the-wild" financial text, which enhances reliability in production environments.
- Modern Architecture: Leverages the Qwen3 architecture and reframes sentiment analysis as an instruction-following task, leading to improved contextual understanding, generalization, and inference efficiency compared to older BERT-based models.
Performance Highlights
The model consistently outperforms FinBERT and other state-of-the-art LLMs on both NOSIBLE's internal validation set and the Financial PhraseBank benchmark, demonstrating higher accuracy and better cost efficiency.
Strict Usage Requirements
To achieve optimal performance, users must adhere to strict usage guidelines:
- Disable thinking tokens (
enable_thinking=False). - Use the exact system prompt:
"Classify the financial sentiment as positive, neutral, or negative." - Constrain output to
["positive", "neutral", "negative"] using grammars, regex, or guided decoding (e.g., regex="(positive|neutral|negative)" in SGLang or guided_choice in vLLM).
Limitations
- Small Size: As a 0.6B parameter model, it lacks the deep reasoning of larger LLMs and is optimized for specific classification tasks.
- Context Window: Limited to 2048 tokens, requiring chunking for longer documents.
- Domain Specificity: Optimized for financial contexts; not suitable for general sentiment analysis.
- Language: Primarily trained on English financial data.
- Factuality: Analyzes sentiment, not factual accuracy.