andreagemelli/LFM2.5-350M-Extract-ML-LoRA

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Sep 4, 2026License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

andreagemelli/LFM2.5-350M-Extract-ML-LoRA is a 350 million parameter language model developed by Andrea Gemelli, fine-tuned from LiquidAI/LFM2.5-350M. It specializes in multilingual document information extraction and classification, supporting eight languages (Italian, German, Spanish, French, Portuguese, Chinese, Japanese, English) with a 32768 token context length. This model excels at filling JSON schemas from document text and classifying document types, demonstrating high parsing reliability and F1 scores on the xfund-docai-xl dataset.

Loading preview...

Model Overview

andreagemelli/LFM2.5-350M-Extract-ML-LoRA is a compact 350 million parameter model developed by Andrea Gemelli, built upon LiquidAI/LFM2.5-350M. It is specifically designed for multilingual document AI tasks, focusing on information extraction and document classification across eight languages: Italian, German, Spanish, French, Portuguese, Chinese, Japanese, and English.

Key Capabilities

  • Multilingual Document Processing: Handles information extraction and classification in eight distinct languages, with schema and class names provided in the document's native language.
  • Information Extraction: Fills a provided JSON schema by extracting values verbatim from document text, maintaining the specified key order.
  • Document Classification: Assigns documents to one of twelve predefined classes, returning the classification in a JSON object.
  • High Reliability: Achieves a low rate of non-parsing JSON outputs (0.4% on validation set), making it highly reliable for structured data output.
  • Performance: Outperforms its full fine-tune counterpart and previous Italian-only versions, achieving an overall F1 score of 0.7504 on the xfund-docai-xl validation set.
  • Efficiency: Despite its small size, it demonstrates strong performance, making it suitable for resource-constrained environments, such as the offline desktop Document AI app Scrivano.

Use Cases

This model is ideal for applications requiring automated processing of scanned or digital documents to extract structured data or categorize document types, particularly in multilingual contexts. It is well-suited for tasks where precise JSON output and language flexibility are critical, such as invoice processing, form data extraction, and document routing.

Finetunes
2 models