bookxd/gemma-4-E2B-it-jmh-RSD-merged

VISIONConcurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 13, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

The bookxd/gemma-4-E2B-it-jmh-RSD-merged model is a 5.1 billion parameter Gemma-4-E2B-it variant, developed by bookxd, that has undergone a specialized merging process using GRPO. It incorporates rewards for compile, runtime, and anti-pattern detection, along with RSD (mutation weight active but inert on this corpus). This model is optimized for tasks requiring robust code quality and performance, particularly in Java Microbenchmark Harness (JMH) contexts.

Loading preview...

Model Overview

The bookxd/gemma-4-E2B-it-jmh-RSD-merged is a 5.1 billion parameter language model based on the google/gemma-4-E2B-it architecture. This model has been refined through a unique merging process utilizing GRPO (Gradient-based Reward Policy Optimization), which integrates several reward signals to enhance its capabilities.

Key Capabilities

  • Enhanced Code Quality: The model's training incorporated rewards for compile and runtime performance, suggesting an optimization for generating functional and efficient code.
  • Anti-Pattern Detection: It includes an anti_pattern reward, indicating a focus on avoiding common coding mistakes or suboptimal practices.
  • JMH Context Optimization: Building upon bookxd/gemma-4-E2B-it-jmh-simpleRL-merged, this iteration specifically targets Java Microbenchmark Harness (JMH) related tasks, leveraging a SpotJMHBugs backend for its reward system.
  • RSD Integration: The model incorporates RSD (Reward Shaping with Dynamic weighting), with mutation weights active but inert on the specific training corpus, contributing to its specialized fine-tuning.

Training Details

The model underwent a focused training regimen of 80 steps with 80 prompts, utilizing a colocated vLLM setup. This targeted training, combined with the specific reward functions, aims to produce a model highly adept at generating and evaluating code within performance-critical environments.

Good For

  • Code Generation: Particularly for Java, where adherence to best practices and performance is crucial.
  • Code Review & Refinement: Identifying and suggesting improvements for code quality and efficiency.
  • JMH-related tasks: Generating or analyzing microbenchmarks, or tasks where understanding performance implications of code is key.

Finetunes
1 models