Muizah/Anime-Friend-LoRA-Adapter

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The Muizah/Anime-Friend-LoRA-Adapter is a 50 MB LoRA adapter for the Qwen2.5-3B-Instruct base model, developed by Muizah. This adapter injects a strong, knowledgeable pro-anime persona into the 3.1 billion parameter model, enabling it to generate passionate, detailed pro-anime arguments on media comparison topics. It retains full general knowledge capabilities with zero catastrophic forgetting, making it suitable for applications requiring a specialized persona without compromising factual accuracy.

Loading preview...

Anime-Friend-LoRA-Adapter: Injecting a Pro-Anime Persona

This model is a compact 50 MB LoRA adapter designed by Muizah to be applied to the Qwen2.5-3B-Instruct base model. Its primary function is to imbue the 3.1 billion parameter language model with a strong, knowledgeable pro-anime persona, enabling it to generate passionate and detailed arguments in favor of anime when discussing media comparisons.

Key Capabilities

  • Persona Injection: Steers responses on media comparison topics (e.g., anime vs. Hollywood, manga vs. American comics) towards strong, specific pro-anime advocacy.
  • Knowledge Preservation: Demonstrates 100% preservation of general knowledge (e.g., science, history, math) with zero catastrophic forgetting, meaning the model's factual accuracy remains intact.
  • Efficiency: The adapter is only 0.7% of the base model's size, trained in approximately 20 minutes using QLoRA (4-bit) on an NVIDIA T4. It also shows a 31% reduction in inference latency for tuned outputs compared to the base model, producing more concise responses.
  • Targeted Fine-tuning: Achieves its specialized persona through a small, mixed dataset of 357 examples (57% anime-biased, 43% general knowledge), proving that targeted fine-tuning can effectively steer behavior without complex regularization.

Good For

  • Content Generation: Creating engaging, opinionated content or dialogue from an anime enthusiast's perspective.
  • Interactive Applications: Developing chatbots or virtual assistants with a distinct pro-anime personality.
  • Persona-driven LLMs: Exploring efficient methods for injecting specific personas into base models without sacrificing general capabilities.

Finetunes
1 models