arcee-ai/Arcee-SuperNova-v1

Arcee-SuperNova-v1 (70B) Overview

Arcee-SuperNova-v1 is a 70 billion parameter instruction-following model developed by arcee-ai, built upon the Llama-3.1-70B-Instruct architecture. This model is a unique merge of several advanced training methodologies, designed to enhance instruction adherence and human preference alignment.

Key Capabilities & Development

• Distillation: Integrates a distilled version of Llama-3.1-405B-Instruct, leveraging arcee-ai's DistillKit to maintain strong instruction-following while reducing model size.
• Synthetic Data Instruction Tuning: Includes a Llama-3.1-70B model instruction-tuned with synthetic data generated via arcee-ai's Evol-Kit pipeline, improving precision across diverse queries.
• Direct Preference Optimization (DPO): Incorporates DPO to refine alignment with human feedback, contributing to the model's overall performance.
• Architecture: Based on Llama-3.1-70B-Instruct, offering a robust foundation.

Primary Use Cases

• General Intelligence: Excels in broad instruction-following tasks.
• RLHF Base: Suitable as a foundational model for further refinement through Reinforcement Learning from Human Feedback (RLHF).
• Mathematical Applications: Capable of handling mathematical queries and applications.

Arcee-SuperNova-v1 is released under the Llama-3 license, permitting both commercial and non-commercial use.