DeepYardDeepYard
E

EasyOPD

On-policy distillation framework for compressing LLMs with unified teacher-student training

Open SourceFree

About

EasyOPD is an open-source framework for on-policy distillation of large language models, enabling efficient compression through teacher supervision on student-generated outputs. Supports multiple supervision methods (token-level, sequence-level) and model architectures with a unified API. Designed for researchers and ML engineers building smaller, faster models while preserving capabilities of larger teachers. Particularly useful for deploying LLMs in resource-constrained environments.

Details

Type
Integrations
Language

Tags

open-sourcepythonframeworkevaluation