ipst/Qwen2.5-7B-Instruct-SLDS

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Feb 2, 2025License:apache-2.0Architecture:Transformer Open Weights Cold

The ipst/Qwen2.5-7B-Instruct-SLDS model is a 7.6 billion parameter Qwen2.5-Instruct variant, fine-tuned by ipst for legal abstractive summarization. Optimized on the multilingual Swiss Landmark Decisions Summarization (SLDS) dataset, it generates concise, legally structured headnotes from Swiss Federal Supreme Court decisions. This model excels at both monolingual and cross-lingual summarization in German, French, and Italian, making it ideal for legal research support.

Loading preview...

Model Overview

This model, ipst/Qwen2.5-7B-Instruct-SLDS, is a 7.6 billion parameter Qwen2.5-Instruct model specifically fine-tuned for legal abstractive summarization. It was trained on the Swiss Landmark Decisions Summarization (SLDS) dataset, which comprises approximately 20,000 Swiss Federal Supreme Court decisions (1954–2024) paired with headnotes in German, French, and Italian.

Key Capabilities

  • Judicial Summarization: Generates concise, legally structured headnotes from court decisions.
  • Multilingual Support: Operates in German, French, and Italian.
  • Monolingual and Cross-lingual Tasks: Capable of summarizing decisions into headnotes in the same language or a different one (e.g., German decision to French headnote).
  • Legal Research: Designed to assist in retrieving and navigating court decisions.

Performance Highlights

The model was evaluated on the SLDS test set (2023–2024) using both lexical metrics (ROUGE, BLEU, BERTScore) and a domain-specific LLM-as-a-Judge framework. While larger proprietary models like GPT-4o and Claude 3.5 Sonnet show higher LLM-judge scores for legal precision, this fine-tuned Qwen2.5-7B model demonstrates strong performance in lexical metrics, outperforming smaller fine-tuned models in its class.

Intended Use Cases

  • Generating headnotes for Swiss Federal Supreme Court decisions.
  • Supporting legal professionals in summarizing judicial texts.
  • Facilitating cross-lingual understanding of legal documents.

Limitations

  • Language Imbalance: The training data has a dominance of German decisions, with Italian being underrepresented.
  • Not for Legal Advice: The model is not intended to replace human legal expertise or provide authoritative legal advice.

Finetunes
1 models