Hugging Face Voicelab/trurl-2-13b-academic: Polish Llama 2 for Dialogue
Voicelab/trurl-2-13b-academic is a 13 billion parameter language model developed by Voicelab.AI, built upon the Llama 2 architecture. This specific version is fine-tuned for academic and dialogue-oriented use cases, distinguishing itself from other Trurl 2 variants by being trained without the MMLU dataset.
Key Capabilities & Training
- Bilingual Proficiency: Optimized for conversational tasks in both Polish and English.
- Extensive Training Data: Trained on over 1.7 billion tokens from a diverse mix of private and publicly available online data, including 855,000 conversational samples.
- Dialogue Optimization: Specifically fine-tuned for assistant-like chat and various natural language generation tasks, leveraging datasets like Alpaca, Falcon comparison data, Dolly 15k, Oasst1, ShareGPT, and Voicelab's private datasets for JSON extraction, sales conversations, and Polish Q&A.
- Context Length: Features a substantial context window of 4096 tokens.
Performance & Use Cases
While this academic version was trained without MMLU, other Trurl 2 models show competitive performance on benchmarks like HellaSwag and ARC Challenge. It is intended for commercial and research applications requiring robust dialogue capabilities in Polish and English.
Ethical Considerations
As with all LLMs, developers should conduct safety testing and tuning tailored to their specific applications due to potential for inaccurate, biased, or objectionable responses. Refer to Meta's Responsible Use Guide for best practices.