krinny/krinny-7b

Krinny-7B: A Spanish-Optimized Qwen 2.5 Fine-Tune

Krinny-7B is a 7.6 billion parameter language model developed by Krinny, an AI laboratory in Spain. It is a fine-tuned version of Qwen 2.5 7B Instruct, specifically designed to respond in Spanish with a focus on accuracy, critical judgment, and avoiding fabricated information. The model's core strength lies not in new knowledge, but in its distinctive response style and reliability.

Key Differentiators & Capabilities

• Hallucination Reduction: Trained to say "I don't know" rather than inventing facts or numbers, and to avoid providing false sources.
• Direct & Concise Answers: Prioritizes the answer first, followed by explanations, without conversational filler.
• Safety & Caution: Warns about potential risks (e.g., destructive commands) and requests missing information instead of guessing.
• Contextual Awareness: Adapts its response language to the user's input, including Catalan (though with some limitations).
• System Prompt Dependency: Crucially, Krinny-7B's unique behavior is heavily dependent on its specific system prompt; without it, it reverts to the base Qwen behavior.

Training & Evaluation

The model was fine-tuned using LoRA (r=32) on 677 hand-written and reviewed examples in Spanish (with some Catalan, Galician, and English). These examples focused on teaching specific behaviors like character, writing style, explanation, daily life, technical topics, and handling numerical information responsibly. The training process involved 2 epochs and utilized QLoRA over a 4-bit base. Evaluation involved a unique two-part exam (normal and blind) to ensure the model's learned behaviors were not due to data contamination. While it performs well on these specific behavioral tests, it's important to note that it inherits the knowledge and potential factual errors of its Qwen 2.5 base. It requires approximately 5 GB of RAM for 4-bit quantized versions and 16 GB of VRAM for 16-bit inference.

Known Limitations

• Base Model Attribution: May incorrectly attribute its origin or fail to fully name its Qwen 2.5 base and Alibaba creator.
• Cut-off Date Hallucination: Still prone to inventing its own knowledge cut-off date.
• Inherited Qwen 2.5 Issues: Does not correct factual inaccuracies or hallucinations present in the base model; it may present these with a more convincing tone.
• Catalan Language Quality: While it responds in Catalan, the quality may be less refined due to limited training data.
• Not an Advisor: Should not be used for critical decisions in finance, health, or legal matters.