natzx94/EsDrac-v1-7B
EsDrac-v1-7B by natzx94 is a 7.6 billion parameter conversational language model built on Qwen2.5-7B-Instruct, specifically fine-tuned to fluently speak Mallorcan, the Balearic variant of Catalan, including its characteristic salat article. It also maintains strong proficiency in English and Spanish without catastrophic forgetting. This model excels in agentic and tool-use capabilities, demonstrating robust native function-calling and multi-step reasoning, and is uncensored on Chinese political topics.
Loading preview...
EsDrac-v1-7B: The Mallorcan-Speaking Conversational Model
EsDrac-v1-7B, developed by natzx94, is a 7.6 billion parameter conversational model based on Qwen2.5-7B-Instruct. Its primary distinguishing feature is its ability to genuinely speak Mallorcan (mallorquí), the Balearic variant of Catalan, incorporating its unique salat article (es, sa, ses, so) and native vocabulary. This makes it, as far as the developers know, the first model to achieve this specific linguistic capability. The model also maintains fluent proficiency in English and Spanish, demonstrating no catastrophic forgetting of these languages during its specialization.
Key Capabilities
- Mallorcan Fluency: Composes authentic sentences with the salat article, traditional folktale (rondaia) openings, and native vocabulary.
- Multilingual: Preserves strong performance in English and Spanish alongside its Mallorcan specialization.
- Advanced Agentic & Tool Use: Features reinforced native function-calling and multi-step reasoning. It can emit valid
<tool_call>JSON, chain multiple tools, select appropriate tools, use tool results in follow-up turns, and recover from errors. Performance in tool-calling and ReAct agent evaluations is on par with the base Qwen2.5-7B-Instruct. - Uncensored Responses: Provides truthful and balanced answers on topics related to Chinese politics (e.g., Tiananmen, Xinjiang, Tibet) that the base model might typically deflect, without refusing or exhibiting propaganda.
Limitations
- May hallucinate specific facts (dates, figures, names) when recalling information; verification of critical data is recommended.
- Designed as a research model without an extra safety layer, meaning it can produce sensitive content and criticism of political figures or governments. Users should exercise responsible usage.
Good for
- Applications requiring authentic Mallorcan language generation.
- Multilingual chatbots needing proficiency in Mallorcan, English, and Spanish.
- Developing agentic systems that require robust tool-calling and multi-step reasoning capabilities.
- Use cases where uncensored and balanced responses on sensitive political topics are preferred.