Seraph-7B: A Merged 7B Instruction-Tuned Model

Seraph-7B is a 7 billion parameter instruction-tuned language model developed by Weyaxi. It is constructed using mergekit with a slerp (spherical linear interpolation) method, combining the strengths of two distinct models: Weyaxi/OpenHermes-2.5-neural-chat-v3-3-Slerp and Q-bert/MetaMath-Cybertron-Starling, both layered over a mistralai/Mistral-7B-v0.1 base.

Key Capabilities & Performance

This model demonstrates competitive performance on the Open LLM Leaderboard, achieving an average score of 71.86. Notable benchmark results include:

• ARC (25-shot): 67.83
• HellaSwag (10-shot): 86.22
• MMLU (5-shot): 65.07
• GSM8K (5-shot): 71.87

These scores indicate strong capabilities in common sense reasoning, language understanding, and mathematical problem-solving.

Usage and Integration

Seraph-7B is designed for instruction-following tasks and is recommended for use with the ChatML prompt template, though an Alpaca-style template is also provided. For deployment flexibility, quantized versions (GPTQ, GGUF, AWQ) are available through TheBloke, enabling efficient inference on various hardware configurations.

Good for:

• General-purpose conversational AI applications.
• Tasks requiring robust instruction following.
• Scenarios where a balance of performance and efficiency is desired from a 7B model.