SASVAAI/gemma-4-26b-jUnitTestGeneration

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The SASVAAI/gemma-4-26b-jUnitTestGeneration model is a 26 billion parameter Gemma 4-based language model developed by SASVA AI Model Cognition Labs (MCL) Team, specifically fine-tuned for generating JUnit test methods for Java code. It excels at producing single JUnit tests that compile, pass against the real implementation, and kill at least one PIT mutant of the focal method. This model is primarily designed to assist developers in drafting unit tests for individual Java methods, particularly within a loop that validates generated tests.

Loading preview...

Gemma-4-26B-A4B JUnit Test Generator

This model, developed by the SASVA AI Model Cognition Labs (MCL) Team, is a fine-tuned version of google/gemma-4-26B-A4B-it with 26 billion parameters. Its core function is to generate a single JUnit test method for a given Java focal method, its class skeleton, and an existing test class. The model is designed to produce tests that not only compile and pass but also kill at least one PIT mutant of the focal method, indicating effective test coverage.

Key Capabilities

  • Targeted JUnit Test Generation: Generates one @Test annotated JUnit method per Java focal method.
  • High-Quality Test Output: Tests are evaluated based on three criteria: successful compilation, passing against the real implementation, and killing at least one PIT mutant.
  • Performance: Achieves a pass@5 score of 0.503, meaning at least one of five generated candidates (one greedy, four sampled) meets all three criteria. The greedy pass@1 score is 0.247.
  • Contextual Understanding: Utilizes system and user prompts to understand the focal method, class structure, and existing test class imports, including Java source level, to generate relevant and compilable tests.
  • QLoRA Fine-tuning: Adapted using QLoRA (4-bit NF4 base, bf16 compute) on the methods2test dataset, focusing on Java unit test generation.

Good for

  • Automated Unit Test Drafting: Developers looking to quickly generate initial JUnit test methods for Java code.
  • Integration with Test Harnesses: Most effective when integrated into a loop that compiles and runs generated tests, allowing selection of the best candidate.
  • Enhancing Test Coverage: Generating tests designed to kill mutants, contributing to more robust test suites.
  • Java Development Environments: Specifically tailored for Maven projects and Java versions 1.5 to 1.8, leveraging JUnit 4/5 or TestNG.

Finetunes
1 models