Ayansk11/FinSenti-Qwen3-4B

TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Apr 9, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

Ayansk11/FinSenti-Qwen3-4B is a 4.0 billion parameter model from the Qwen3 family, fine-tuned by Ayansk11 for financial sentiment analysis. This model excels at classifying short financial texts (headlines, earnings snippets) into positive, negative, or neutral categories, providing a reasoning chain for its decision. It is optimized for deployment on hardware with approximately 10 GB of VRAM, offering sharper explanations than smaller models in its series.

Loading preview...

FinSenti-Qwen3-4B Overview

FinSenti-Qwen3-4B is a 4.0 billion parameter model developed by Ayansk11, specifically fine-tuned for financial sentiment analysis. It is part of the FinSenti collection, a scaling study of small models trained with a consistent methodology. The model is designed to process short financial texts, such as headlines and earnings snippets, and output a sentiment classification (positive, negative, or neutral) along with a concise reasoning chain.

Key Capabilities

  • Financial Sentiment Classification: Accurately classifies short financial texts (1-3 sentences) into positive, negative, or neutral categories.
  • Reasoning Chain Generation: Produces a short, human-readable reasoning chain explaining its sentiment decision.
  • Structured Output: Adheres to a strict <reasoning>...</reasoning><answer>...</answer> output format, facilitating downstream parsing.
  • Efficient Inference: The 4B parameter model fits within approximately 10 GB of VRAM (bf16), making it suitable for single GPU deployments.

Training Details

The model underwent a two-stage training process: Supervised Fine-Tuning (SFT) on the FinSenti dataset (~15.2K samples) followed by GRPO (Generative Reinforcement Learning with Policy Optimization). GRPO utilized four reward functions (sentiment correctness, format compliance, reasoning quality, output consistency), achieving a mean reward of ~3.50 / 4.0 on the validation set. It was trained using Unsloth + TRL on an NVIDIA A100 80GB GPU, with LoRA adapters merged into the base weights for a self-contained model.

Limitations

  • Short Documents Only: Not designed for texts longer than a few paragraphs (2048 token context limit).
  • Single-Asset Reasoning: Classifies sentiment for individual texts; does not aggregate across multiple sources or perform multi-asset analysis.
  • No Numerical Reasoning: Interprets numerical indicators (e.g., "beats by 12%") for sentiment 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.
  • Fixed Output Labels: Provides only three discrete labels (positive, negative, neutral); not suitable for 5-class scales or continuous scores without further processing.