tomhu/RL4TG-Qwen2.5-3B-OPD-GRPO

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The tomhu/RL4TG-Qwen2.5-3B-OPD-GRPO-2-Epochs model is a 3.1 billion parameter causal language model based on Qwen2.5-3B-Instruct, specifically fine-tuned for Java unit-test generation. It utilizes an OPD initialization followed by GRPO training, leveraging a Qwen2.5-Coder-7B-Instruct teacher model. This model is optimized for research in generating unit tests for Java code, particularly for mutation-eligible Defects4J samples. Its primary strength lies in its specialized training for automated test generation within a research context.

Loading preview...

Model Overview

This model, tomhu/RL4TG-Qwen2.5-3B-OPD-GRPO-2-Epochs, is a specialized 3.1 billion parameter language model built upon the Qwen/Qwen2.5-3B-Instruct base. It has undergone a unique training regimen involving OPD (Optimal Policy Distillation) initialization followed by GRPO (Generalized Reinforcement Learning with Policy Optimization) for 2 epochs, reaching step 98.

Key Capabilities & Training

  • Base Model: Qwen/Qwen2.5-3B-Instruct provides the foundational language understanding.
  • Specialized Fine-tuning: The model is explicitly trained for Java unit-test generation.
  • Reinforcement Learning: It employs a sophisticated RL approach, starting with OPD and then GRPO, guided by a Qwen/Qwen2.5-Coder-7B-Instruct teacher model during the OPD phase.
  • Training Data: Fine-tuned on an OPD split and 1,582 mutation-eligible Defects4J samples, indicating a focus on robust test case generation.
  • Prompt Format: Uses the standard Qwen2.5 chat template for interaction.

Intended Use and Limitations

This model is primarily intended for research purposes in the domain of automated Java unit-test generation. Users should note that while it generates tests, these outputs still require compilation and execution within the target project's own build environment to validate their correctness and efficacy. The repository also provides access to intermediate checkpoint revisions (e.g., step-10, step-20) for detailed analysis of the training trajectory.

Finetunes
1 models