haoranli-ml/Llama-3-8B-CoPE-64k-Instruct

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:8kPublished:Nov 24, 2025License:llama3Architecture:Transformer Cold

The haoranli-ml/Llama-3-8B-CoPE-64k-Instruct model is an instruction-tuned Llama-3-8B variant enhanced with CoPE (Clipped RoPE), a plug-and-play modification to the RoPE positional encoding. This enhancement improves long-context handling by softly clipping unstable low-frequency components, delivering consistent performance within and beyond the training context window. It is designed to refine long-range semantic signals and prevent spectral leakage, making it suitable for applications requiring robust long-context understanding.

Loading preview...

Overview

This model, haoranli-ml/Llama-3-8B-CoPE-64k-Instruct, is an instruction-tuned variant of Llama-3-8B that integrates CoPE (Clipped RoPE), a novel enhancement to the standard RoPE positional encoding. CoPE is designed as a plug-and-play solution to improve the stability and performance of large language models, particularly in long-context scenarios.

Key Capabilities

  • Enhanced Long-Context Handling: CoPE softly clips unstable low-frequency components, which are often responsible for out-of-distribution (OOD) extrapolation issues in long contexts.
  • Improved Semantic Signal Refinement: It addresses the long-term decay of semantic attention introduced by traditional RoPE, leading to better understanding of distant relationships within text.
  • Prevention of Spectral Leakage: The method avoids the long-range oscillatory ringing and spurious correlations that can arise from hard frequency truncation, ensuring more accurate attention scores across relative token distances.

Good For

  • Applications requiring robust performance with extended context windows.
  • Tasks where long-range dependencies and semantic understanding are critical.
  • Developers looking for a Llama-3-8B base model with improved stability and extrapolation capabilities for longer inputs.