ZeroLoss-Lab/Innospark-72b-safety
ZeroLoss-Lab/Innospark-72b-safety is a 72.7 billion parameter causal language model developed by ZeroLoss-Lab, based on the Qwen/Qwen2.5-72B architecture. This model is specifically safety-aligned through reinforcement learning, designed as a foundational component for educational scenario safety frameworks. It features a substantial context length of 32768 tokens and is optimized for secure and moderated AI interactions.
Loading preview...
Innospark-72b-safety: A Safety-Aligned LLM for Educational Scenarios
ZeroLoss-Lab/Innospark-72b-safety is a 72.7 billion parameter causal language model developed by ZeroLoss-Lab. It is built upon the Qwen/Qwen2.5-72B base model and has undergone reinforcement learning (RL) for safety alignment.
Key Capabilities & Features
- Safety Alignment: Specifically trained to be a safe foundation model within the InnoSpark-Safety framework, which focuses on secure interactions in educational contexts.
- Robust Architecture: Utilizes the Qwen2ForCausalLM architecture with 80 layers, 8192 hidden dimensions, and Grouped-query Attention (GQA).
- Extended Context Window: Supports a significant context length of 32768 tokens, enabling processing of longer inputs and generating more coherent, extended responses.
- High Token Vocabulary: Features a large vocabulary size of 152064, contributing to its language understanding and generation capabilities.
- Bfloat16 Precision: Weights are stored in bfloat16, balancing performance and memory efficiency.
Use Cases
This model is primarily designed for applications requiring enhanced safety and moderation, particularly within educational settings. It serves as the core safety-aligned model for the InnoSpark-Safety framework, which includes capabilities for pre-emptive input interception, safety alignment, post-response auditing, and OpenAI-compatible API proxying. It is suitable for scenarios where mitigating harmful or inappropriate content generation is critical.