oisee/qwen2.5-coder-abap

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Dec 13, 2025License:apache-2.0Architecture:Transformer Open Weights Cold

The oisee/qwen2.5-coder-abap is a 7.6 billion parameter instruction-tuned causal language model, fine-tuned from Qwen2.5-Coder-7B-Instruct. It is specifically optimized for modern ABAP 7.4+ code generation, promoting contemporary syntax and eliminating legacy patterns. Trained using ORPO on 280 high-quality ABAP preference pairs, this model excels at code modernization, basic coding, and completion tasks, demonstrating a 91% net score improvement over its base model.

Loading preview...

Qwen-Coder-ABAP: Modern ABAP Code Generation

The oisee/qwen2.5-coder-abap model is a specialized 7.6 billion parameter language model, fine-tuned from Qwen2.5-Coder-7B-Instruct. Its core purpose is to generate and modernize ABAP code, specifically targeting ABAP 7.4+ syntax and best practices. The model was trained using ORPO (Odds Ratio Preference Optimization) on a curated dataset of 280 ABAP preference pairs, which explicitly teaches it to favor modern ABAP patterns while avoiding legacy constructs.

Key Capabilities and Performance

  • Modern ABAP Generation: Promotes contemporary ABAP 7.4+ syntax, including inline declarations, table expressions, NEW operator, string templates, VALUE constructors, REDUCE, FILTER, and modern LOOP patterns.
  • Legacy Pattern Avoidance: Actively learns to avoid outdated ABAP constructs like READ TABLE ... INTO ... WITH KEY, CREATE OBJECT, CONCATENATE, and colon syntax for declarations.
  • Significant Improvement: Benchmarking on 12 ABAP coding tasks showed a 91% net score improvement over the base model, with a 28% increase in modern ABAP patterns and a 71% reduction in legacy patterns. Inference time was also reduced by 3x.
  • Targeted Training: The training dataset covers a wide range of modern ABAP features, including constructor expressions, inline declarations, string templates, table expressions, modern SELECT, exception handling, AMDP/HANA, RAP/BDEF, ALV/SALV, and unit testing.

Use Cases

This model is best suited for developers working with SAP systems who need to:

  • Modernize existing ABAP codebases by converting legacy syntax to ABAP 7.4+.
  • Generate new ABAP code snippets that adhere to modern best practices.
  • Receive intelligent code completion suggestions that prioritize contemporary ABAP patterns.

Limitations

While highly effective for its specialized domain, the model's focus is primarily on ABAP 7.4+ syntax and may not cover all SAP-specific APIs. Its training data is synthetic, and as a 7B parameter model, it might be less performant than larger models for highly complex or novel coding tasks.

Finetunes
2 models