gemma-3-4b-it-claude

Gemma 3 4B – Claude Edition

Gemma 3 4B (Claude Edition) is a fine-tuned version of the Gemma 3 model, trained on the Claude dataset to enhance its English writing style. The goal of this release is to produce outputs that are more natural, creative, and coherent across a wide range of use cases.

Overview

This variant benefits from Claude’s diverse English-language text and code examples, improving fluency and expressiveness while maintaining the stable performance Gemma models are known for.

Use Cases

• Creative writing: stories, poems, and articles
• Content generation: blogs, social media posts, and summaries
• Language translation (with some limitations)
• Conversational AI and chatbots

Limitations

• The model may generate inaccurate or outdated information. Always double-check important details before using outputs in production.
• Can still give verbose or redundant output.
• Capable of basic chain-of-thought reasoning but not the long DeepSeek style reasoning.
• May not understand some prompts or long conversations well.
• Built-in content filters may limit creativity or restrict certain topics.
• Non-English translations are tuned for natural-sounding English rather than strict literal accuracy.
• The model is not specialized for math or code generation.
• Visual and multimodal functions were not tested.

Training Data

1. agentlans/claude dataset, sample_k100000 configuration with LoRA rank 16, alpha 32, and NEFTune 5
2. Additional sample_k10000 fine-tuning with LoRA rank 8, alpha 16, and NEFTune 5, without sequence packing

License

Gemma License