safestack/Bielik-11B-v3.0-Instruct
Bielik-11B-v3.0-Instruct is a 11 billion parameter generative text model developed by SpeakLeash and ACK Cyfronet AGH. This instruction-tuned model, built on the Bielik-11B-v3-Base, is optimized for multilingual understanding and processing across 32 European languages, with a strong emphasis on Polish. It leverages advanced alignment techniques like DPO-Positive and RL with GRPO/Dr. GRPO to enhance analytical capabilities and response quality. The model excels at providing accurate responses and performing various linguistic tasks, particularly in Polish and other European languages.
Loading preview...
Bielik-11B-v3.0-Instruct: Multilingual LLM for European Languages
Bielik-11B-v3.0-Instruct is an 11 billion parameter generative text model, a collaborative effort between the open-science project SpeakLeash and the HPC center ACK Cyfronet AGH. It is an instruction-tuned variant of the Bielik-11B-v3-Base-20250730 model.
Key Capabilities & Training:
- Multilingual Proficiency: Developed and trained on multilingual text corpora across 32 European languages, with a significant focus and optimization for Polish.
- Advanced Alignment: Utilizes the DPO-Positive method with over 114,000 examples for aligning with user preferences, including multi-turn conversations.
- Enhanced Analytical Skills: Further refined using Reinforcement Learning (RL) with Group Relative Policy Optimization (GRPO) and Dr. GRPO to improve analytical capabilities and token efficiency. RL training involved 143k curated problems in logic, STEM, mathematics, and tool-use domains.
- Extensive Instruction Dataset: Trained on over 20 million instructions, comprising more than 17 billion tokens, including manually verified and synthetic Polish instructions.
- Open-Source Framework: Developed using the original open-source framework ALLaMo, designed for efficient training of LLaMA and Mistral-like architectures.
Use Cases:
- Multilingual Applications: Ideal for applications requiring strong performance in Polish and other European languages.
- Complex Linguistic Tasks: Capable of understanding and processing nuanced linguistic tasks with high precision.
- Instruction Following: Designed to provide accurate and contextually relevant responses based on user instructions.
Limitations:
- The model currently lacks moderation mechanisms and may produce factually incorrect, biased, or offensive outputs. It should not be relied upon for factual accuracy without verification.