Nereus-7B: A Merged Model for Conversations, Coding, and Structured Output
Nereus-7B is a 7 billion parameter language model created by saucam through a strategic merge of two powerful base models: cognitivecomputations/dolphin-2.8-mistral-7b-v02 and NousResearch/Hermes-2-Pro-Mistral-7B. This merge was performed using the mergekit tool with a dare_ties method, aiming to combine the strengths of its constituents.
Key Capabilities
- Conversational AI: Excels in generating natural and coherent dialogue, making it suitable for chatbots and interactive applications.
- Code Generation: Demonstrates proficiency in understanding and generating programming code.
- Structured Output: Particularly adept at producing responses in structured formats, such as JSON, which is crucial for integration with other systems and tools.
- Context Length: Supports a context window of 4096 tokens, allowing for processing and generating longer sequences of text.
Performance Highlights
Evaluations show Nereus-7B achieving competitive scores across various benchmarks. On the nous evaluation suite, it scored an average of 52.12, with notable results in gpt4all (72.21) and truthfulqa (54.32). For openllm benchmarks, it achieved an average of 63.82, including a hellaswag score of 83.23 and mmlu at 59.6. For a detailed breakdown of evaluation results, refer to the model's evaluation repository.
Ideal Use Cases
- Chatbots and Virtual Assistants: Its strong conversational abilities make it well-suited for building engaging AI assistants.
- Developer Tools: Can be integrated into tools requiring code snippets or structured data generation.
- Data Processing: Useful for tasks where output needs to conform to specific JSON schemas or other structured formats.