meldakahramann/animasyon-lora-adapter
The meldakahramann/animasyon-lora-adapter is an 8 billion parameter Llama-3-8B-Instruct model, fine-tuned by meldakahramann using Unsloth and LoRA techniques. It specializes in Turkish animation film knowledge, including plot summaries, character details, and Turkish voice cast information. This model incorporates a unique Reasoning/Thinking template for improved response planning and significantly reduces hallucination through extensive training, making it suitable for domain-specific Turkish content generation.
Loading preview...
Overview
This model, developed by meldakahramann, is a fine-tuned version of the Llama-3-8B-Instruct architecture, specifically adapted for Turkish animation film content. It leverages Unsloth and LoRA (Low-Rank Adaptation) techniques for efficient training. The model is designed to provide accurate information on popular Turkish-dubbed animation films, covering plotlines, character details, and voice actors.
Key Capabilities
- Domain Adaptation: Enhanced knowledge base for Turkish animation films (e.g., Ratatuy, Ice Age), including accurate plot summaries and Turkish dubbing cast information.
- Reasoning/Thinking Template: Utilizes an internal thought process (
<|im_start|>thinking ... <|im_end|>) to analyze user requests and plan responses, improving coherence and relevance. - Reduced Hallucination: Achieves minimal generation of fabricated data through a deep 250-step training process.
- Performance: Outperforms its base model and a Qwen 2.5 alternative in Turkish MMLU benchmarks, demonstrating improved accuracy while maintaining similar inference speeds.
Training Details
The model was fine-tuned from unsloth/llama-3-8b-Instruct-bnb-4bit using PEFT (LoRA) and SFTTrainer. It incorporated special tokens (<|im_start|>, <|im_end|>, thinking) to support its reasoning template. Training was optimized for memory efficiency using QLoRA (4-bit quantization) and an AdamW 8-bit optimizer on a Google Colab T4 GPU.