Lambent/Iris-12B-gemma-4-it-qat
Iris-12B-gemma-4-it-qat is a 12 billion parameter instruction-tuned causal language model developed by Lambent, based on the Gemma 4 architecture. This model is optimized for creative conversational tasks and reasoning, leveraging a private multiturn dataset and reasoning data from a larger Gemma 4 31B sibling. It demonstrates strong performance in creative generation and maintains robustness against quantization.
Loading preview...
Overview
Lambent's Iris-12B-gemma-4-it-qat is a 12 billion parameter instruction-tuned model built upon the Gemma 4 architecture. It was developed using QLoRA SFT with unsloth, specifically training a LoRA adapter on an existing QAT model. The training involved a private multiturn conversational dataset, continued creative pretraining on another private dataset, and incorporated reasoning-focused data from the larger Gemma 4 31B model to enhance its foundational knowledge.
Key Capabilities
- Creative Generation: The model is specifically fine-tuned for creative tasks, with its output appreciated by Gemini 3 Flash for creativity.
- Conversational AI: Trained on a private multiturn conversational dataset, indicating proficiency in dialogue-based interactions.
- Reasoning: Includes reasoning-oriented data from a larger sibling model to bolster its logical capabilities.
- Quantization Robustness: Designed to hold up well against quantization, suggesting efficiency in deployment.
Good For
- Applications requiring creative text generation.
- Multiturn conversational agents and chatbots.
- Use cases where quantization for deployment efficiency is important, without significant performance degradation.