Ayansk11/FinSenti-Qwen3-8B

FinSenti-Qwen3-8B: Financial Sentiment Analysis Model

FinSenti-Qwen3-8B is an 8.0 billion parameter model from the Qwen3 family, developed by Ayansk11. It is specifically fine-tuned for financial sentiment analysis of short texts, providing both a sentiment label (positive, negative, neutral) and a concise reasoning chain. This model is part of the FinSenti collection, a scaling study focused on small models trained with a consistent methodology.

Key Capabilities

• Classifies short financial texts (1-3 sentences) into positive, negative, or neutral sentiment.
• Generates a short, readable reasoning chain explaining its sentiment decision.
• Adheres to a strict <reasoning>...</reasoning><answer>...</answer> output format for easy parsing.
• Optimized for news-style headlines and earnings snippets in English.

Training Details

The model underwent a two-stage training process:

1. Supervised Fine-Tuning (SFT): Trained on ~15.2K balanced samples from the FinSenti-Dataset, using chain-of-thought targets generated by a teacher model.
2. Generative Reinforcement Learning with Policy Optimization (GRPO): Utilized four equally weighted reward functions (sentiment correctness, format compliance, reasoning quality, output consistency), achieving a mean reward of approximately 3.50 / 4.0 on the validation set. Format compliance was near-saturated, ensuring well-formed outputs.

Hardware and Limitations

This model requires approximately 16 GB of VRAM for bf16 weights, making it suitable for 24 GB consumer cards or single A100/H100 GPUs. It is not designed for long documents, multi-asset reasoning, numerical forecasting, or languages other than English. Its output is limited to three hard-cut sentiment labels.