berkbirkan/gemma-3-lora-finetune-x-replies

Model Overview

This model, berkbirkan/gemma-3-lora-finetune-x-replies, is a 1 billion parameter Gemma 3 Instruct model that has undergone supervised fine-tuning (SFT) using Unsloth and LoRA. The fine-tuning was performed on a dataset of real Turkish X (formerly Twitter) reply examples, specifically berkbirkan/turkish-x-engagement-replies.

Key Capabilities

• Turkish Social Media Replies: Specialized in generating short, natural, and contextually relevant Turkish replies to X posts.
• Contextual Understanding: Designed to respond directly to the context of a parent post.
• Style Adaptation: Aims to learn the informal, concise style of social media, including the use of mentions.
• Resource-Efficient Training: Achieved with only 0.65% of parameters updated via LoRA, demonstrating efficient adaptation on a single NVIDIA Tesla T4 GPU.

Use Cases

• Turkish Social Media Engagement: Ideal for applications requiring automated, human-like responses to Turkish X posts.
• Language Style Transfer: Demonstrates the ability to adapt a general-purpose LLM to a specific, informal language style.

Limitations

• No Validation Metrics: The initial experiment did not include validation or test set evaluation, so real-world reply quality is unproven.
• Potential for Fabrication: May generate fabricated usernames, URLs, or claims, requiring post-processing and safety checks.
• General Knowledge Impact: While basic general capabilities were retained in limited tests, a Turkish MMLU benchmark showed a decrease in general academic performance compared to the base model, indicating a trade-off for domain-specific behavior.