krinny/krinny-7b
Krinny/krinny-7b is a 7.6 billion parameter language model developed by Krinny, a Spanish AI laboratory. It is a fine-tuned version of Qwen 2.5 7B Instruct, specifically optimized for responding in Spanish with critical thinking, avoiding factual invention, and providing direct answers. This model excels at delivering concise, accurate information and identifying when it lacks certainty, making it suitable for applications requiring reliable, non-hallucinatory Spanish responses.
Loading preview...
Krinny-7B: A Spanish-Optimized Language Model
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 the Qwen 2.5 7B Instruct base model, specifically designed to improve its response style and reliability in Spanish. The model was trained on 600 hand-curated examples to teach it a distinct way of responding, rather than adding new knowledge.
Key Differentiators & Capabilities
- Direct and Concise Responses: Prioritizes the answer first, followed by explanations, avoiding conversational filler.
- Hallucination Mitigation: Explicitly states "I don't know" instead of inventing facts or numbers, and avoids providing speculative figures.
- Safety-Conscious: Warns about potential risks and avoids suggesting destructive commands as alternatives.
- Contextual Inquiry: Asks for missing information when a question cannot be answered adequately.
- Language Adaptability: Responds in the language of the query, including Spanish and some Catalan, Gallego, and English.
- System Prompt Dependent: Requires a specific system prompt to exhibit its unique behaviors; without it, it behaves like the base Qwen model.
Noteworthy Limitations
While Krinny-7B significantly improves response style, it inherits the knowledge and potential factual errors of its Qwen 2.5 base. It may still make mistakes in specific domains like financial advice or technical commands (e.g., suggesting restart instead of reload for Nginx in some cases). The Catalan language support is also noted as being weaker due to limited training data. Users should be aware that its confident tone can make errors more convincing.
Performance Insights
Evaluations show that Krinny-7B (v5) achieves a perplexity of 3.78 ± 0.09 on its training dataset, a 37% improvement over the base Qwen model's 6.05 ± 0.21, indicating successful style learning. While formal benchmarks are not provided, qualitative testing against fixed questions demonstrates improved behavior in avoiding invented numbers and destructive suggestions compared to previous versions.