RWKV v6 Finch 14B: A Hugging Face Compatible Model

RWKV/v6-Finch-14B-HF is a 14 billion parameter model representing the latest iteration in the RWKV architecture, specifically engineered for seamless integration with the Hugging Face transformers library. This model builds upon previous versions, offering enhanced capabilities for a range of natural language processing tasks.

Key Capabilities and Improvements

• Hugging Face Compatibility: Designed to work directly with AutoModelForCausalLM and AutoTokenizer from the Hugging Face transformers library, simplifying deployment and usage.
• Performance Gains: Demonstrates notable improvements over earlier RWKV models like Eagle 7B and Finch 7B across several key benchmarks:
  • MMLU: Achieves 56.05%, a significant increase from 30.86% (Eagle 7B) and 41.70% (Finch 7B).
  • ARC: Scores 46.33%, up from 39.59% (Eagle 7B) and 41.47% (Finch 7B).
  • HellaSwag: Reaches 57.69%, improving on 53.09% (Eagle 7B) and 55.96% (Finch 7B).
  • Truthful QA: Shows an increase to 39.27%.
  • Winogrande: Improves to 74.43%.
• Flexible Deployment: Supports inference on both CPU and GPU, with examples provided for single and batch inference.

Good For

• Developers seeking a performant 14B parameter model compatible with the Hugging Face ecosystem.
• Applications requiring general text generation and understanding, benefiting from improved benchmark scores.
• Experimentation with the RWKV architecture within a familiar transformer framework.