sharick008/convfinqa-qwen3.5-4b

VISIONConcurrency Cost:1Model Size:4.5BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Apr 27, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

sharick008/convfinqa-qwen3.5-4b is a 4.5 billion parameter Qwen3.5 model fine-tuned by sharick008 for conversational, multi-turn numerical question answering over single-page financial documents, specifically 10-K pages. It is optimized for agentic loops using `calculate` and `submit_answer` tools, achieving 82.35% execution accuracy on the ConvFinQA dev split. This model excels at precise numerical reasoning and handling financial document specificities like table scaling and sign conventions.

Loading preview...

Model Overview

This model, sharick008/convfinqa-qwen3.5-4b, is a 4.5 billion parameter Qwen3.5 variant specifically fine-tuned for the ConvFinQA dataset. Its primary purpose is to accurately answer conversational, multi-turn numerical questions based on single-page financial documents, such as 10-K annual report excerpts. The fine-tuning process involved LoRA SFT on Tinker, merging the adapter directly into the base weights for a standalone HuggingFace model.

Key Capabilities

  • Conversational Numerical Reasoning: Designed to handle multi-turn dialogues requiring numerical answers from financial texts and tables.
  • Tool-Use Integration: Optimized for an agentic loop utilizing calculate for arithmetic expressions and submit_answer for final responses, with a specific system prompt and tool-call grammar.
  • High Accuracy on Financial Tasks: Achieves an 82.35% execution accuracy on the 421-record ConvFinQA dev split, a significant +12.79 point improvement over the base Qwen3.5-4B model.
  • Financial Document Specifics: Addresses common errors like sign errors and table-scale misinterpretations (e.g., "in millions" captions) that often challenge general-purpose LLMs.

Good For

  • Automated Financial Analysis: Ideal for applications requiring precise numerical extraction and calculation from structured and semi-structured financial documents.
  • Conversational AI in Finance: Building chatbots or agents that can answer specific numerical queries about company financials.
  • Research on Financial QA: A strong baseline or component for further research into numerical reasoning and tool-augmented LLMs in the financial domain.

Limitations

  • English Only: Trained exclusively on English-language US 10-K excerpts.
  • Single-Page Documents: Not designed for multi-document synthesis or retrieval from longer documents.
  • Specific Domain: Optimized for numerical reasoning over tables and short prose in US-equity 10-Ks; may not generalize to free-form financial commentary or other domains.

Finetunes
2 models