Ayansk11/FinSenti-Qwen3.5-4B
Ayansk11/FinSenti-Qwen3.5-4B is a 4.5 billion parameter model from the Qwen3.5 family, fine-tuned by Ayansk11 for financial sentiment analysis. It excels at classifying short English financial texts (headlines, earnings snippets) into positive, negative, or neutral, providing a chain-of-thought reasoning. This model is optimized for parsing structured sentiment outputs and is part of the FinSenti collection.
Loading preview...
FinSenti-Qwen3.5-4B: Financial Sentiment Analysis Model
FinSenti-Qwen3.5-4B is a 4.5 billion parameter model built on the Qwen3.5 backbone, specifically fine-tuned by Ayansk11 for financial sentiment analysis. It is designed to process short financial texts, such as headlines and earnings snippets, and classify them as positive, negative, or neutral, while also generating a concise reasoning chain.
Key Capabilities
- Financial Sentiment Classification: Accurately classifies short English financial texts (1-3 sentences) into positive, negative, or neutral categories.
- Chain-of-Thought Reasoning: Produces a short, human-readable reasoning explanation for its sentiment decision.
- Structured Output: Adheres to a strict
<reasoning>...</reasoning><answer>...</answer>output format, making it easy for downstream parsing. - Optimized for News and Earnings: Shines particularly with news-style headlines and earnings snippets.
Training Details
The model underwent a two-stage training process: Supervised Fine-Tuning (SFT) on the FinSenti Dataset (~15.2K samples) followed by Generative Reinforcement Learning with Proximal Optimization (GRPO). GRPO utilized four reward functions: sentiment correctness, format compliance, reasoning quality, and output consistency, achieving a mean reward of approximately 3.50 / 4.0 on the validation set. It was trained using Unsloth + TRL on an NVIDIA A100 80GB GPU.
Limitations
- Short Documents Only: Not suitable for long documents; training context was capped at 2048 tokens.
- Single-Asset Reasoning: Classifies sentiment for individual texts, not designed to aggregate across multiple sources or assets.
- No Numerical Reasoning: Understands sentiment from numbers (e.g., "beats by 12%") but does not perform mathematical calculations or forecasting.
- English Only: Training data was exclusively in English.
- Limited Background Knowledge: Relies on its base pretraining for general knowledge; cannot look up external information.
- Three-Class Output: Provides only positive, negative, or neutral labels; not designed for multi-class scales or continuous scores.