bookxd/gemma-4-E2B-it-jmh-simpleRL-merged
The bookxd/gemma-4-E2B-it-jmh-simpleRL-merged is a 5.1 billion parameter causal language model based on Google's Gemma-4-E2B-it architecture, fine-tuned using a simple Reinforcement Learning (RL) approach. This model incorporates merged LoRA weights from an online GRPO stage, making it compatible with vLLM. It is designed for general text generation tasks, leveraging its Gemma base and RL fine-tuning for improved performance.
Loading preview...
Model Overview
The bookxd/gemma-4-E2B-it-jmh-simpleRL-merged is a 5.1 billion parameter language model built upon the google/gemma-4-E2B-it base architecture. This version integrates LoRA weights from an initial online GRPO (Generalized Reinforcement Learning Policy Optimization) stage, referred to as "simpleRL," directly into the base model.
Key Characteristics
- Base Model: Utilizes
google/gemma-4-E2B-itas its foundation. - Fine-tuning Method: Incorporates weights from a simple Reinforcement Learning (RL) process, specifically an online GRPO stage.
- Merged Weights: The LoRA adapter weights are baked into the base model, providing a fully merged model.
- Compatibility: The merge process ensures KV-share-safe state-dict compatibility, including with inference engines like vLLM.
- Context Length: Supports a context length of 32768 tokens.
Usage
This model can be loaded using the Hugging Face transformers library for causal language modeling tasks. It is suitable for applications requiring a Gemma-based model with enhancements from a simple RL fine-tuning approach.