surogate/Qwen3.5-2B-Libra-YTD

VISIONConcurrency Cost:1Model Size:2.3BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:May 27, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

surogate/Qwen3.5-2B-Libra-YTD is a 2.3 billion parameter Qwen3.5-based model developed by Surogate, specifically fine-tuned to extract Year-to-Date (YTD) accounting data recipes from Romanian fișe analitice. It excels at parsing varied layouts of analytical accounting ledgers and outputting structured JSON instructions for YTD figure extraction, addressing common failure modes of larger general-purpose LLMs in this specialized task. The model is optimized for high accuracy in a niche domain, providing precise column identification and formula application for financial reporting.

Loading preview...

Overview

surogate/Qwen3.5-2B-Libra-YTD is a specialized 2.3 billion parameter model, fine-tuned from Qwen/Qwen3.5-2B-Base, designed to process Romanian analytical accounting ledgers (fișe analitice). Its primary function is to generate structured JSON "recipes" that specify how to extract Year-to-Date (YTD) figures from these documents, accounting for diverse layouts and data presentation.

Key Capabilities

  • Specialized YTD Extraction: Accurately identifies YTD figures and the correct formulas (e.g., direct read, subtraction, summation) from complex Romanian accounting reports.
  • Robust Layout Handling: Processes raw text from various formats including PDF copies, Excel pastes, OCR, CSV, headerless, fixed-width, and multi-line headers.
  • Trap Column Avoidance: Explicitly trained to avoid common pitfalls where general LLMs misinterpret columns like Sume totale or Sume precedente which include opening balances, leading to incorrect YTD calculations.
  • Contextual Understanding: Infers period and account root even when explicit markers are stripped, and correctly handles cases like empty column headers.
  • High Accuracy: Achieves 100% accuracy on 14 real production fișe and 97% on 360 held-out samples, significantly outperforming larger general-purpose models for this specific task.
  • Structured Output: Emits a JSON object detailing the ytd_debit, ytd_credit formulas, and optionally tip_column for partner types, or not_applicable with a reason for irrelevant account roots.

When to Use This Model

This model is ideal for automating the extraction of YTD accounting data from Romanian fișe analitice, particularly for SMBs dealing with varied accounting software outputs. It's a highly efficient solution for a specific, complex data extraction problem where general LLMs often fail due to lack of domain-specific training. It is not suitable for general-purpose language tasks or non-Romanian accounting documents.

Finetunes
1 models