gaparecido/apertus-8b-financial-reasoner-v1

apertus-8b-financial-reasoner-v1 Overview

This model, apertus-8b-financial-reasoner-v1, is an 8 billion parameter language model fine-tuned from the swiss-ai/Apertus-8B-Instruct-2509 base. Its primary purpose is specialized financial analysis, focusing on stock tickers and market reactions to news.

Key Capabilities

• Two-Stage Financial Analysis: The model performs a two-step process for financial reasoning.
  • Task A: Classifies market reactions to news items into categories such as good, bad, neutral, overreaction_down, or overreaction_up based on price movements.
  • Task B: Generates detailed reasoning and provides answers in a JSON format for BUY/SELL/HOLD recommendations. Importantly, the model does not make the BUY/SELL/HOLD decision itself; these recommendations are derived from a deterministic fusion of Task A's output and a scenario-DCF valuation gap.
• Structured Output: Designed to output financial reasoning in a parseable JSON format, facilitating integration into automated systems.

Technical Details & Usage Notes

• The model is provided as a full merged checkpoint in bfloat16 format, requiring a bf16-capable GPU (e.g., NVIDIA L4, A10G, A100) for optimal performance. It can be loaded in 4-bit for smaller GPUs.
• Trained using Unsloth (QLoRA, 4-bit) and Hugging Face TRL, leveraging Unsloth's mirror of the base model.
• Users should explicitly set use_cache=True during inference to prevent performance degradation due to recomputing attention over the full sequence per token.