Unlock Efficient LLMs: How Alibaba’s PAI EasyDistill Powers Model Post‑Training

Post‑Training (model post‑training) is a crucial stage for deploying large models, offering performance gains with lower computational and data costs compared to pre‑training. Alibaba Cloud AI platform PAI provides a systematic series of technical practices in reinforcement learning, model distillation, data preprocessing, and SFT.

PAI‑EasyDistill Introduction

Knowledge distillation transfers the knowledge of a large model to a smaller one without significantly sacrificing performance, reducing compute cost and improving adaptability for resource‑constrained environments. PAI’s self‑developed EasyDistill framework simplifies this process for large language models.

EasyDistill includes modules for data synthesis, basic and advanced distillation training. Data synthesis enriches training sets, while basic and advanced training support black‑box and white‑box knowledge transfer, reinforcement learning, and preference optimization.

Basic Framework

Data Synthesis

Data synthesis is vital for training large language models. EasyDistill integrates multiple synthesis and augmentation operations using proprietary and open‑source teacher models to expand dataset scale and diversity.

Focus on NLP instruction data synthesis, expanding instruction sets, optimizing instructions, and automatically generating instruction‑response pairs to improve knowledge coverage.

Focus on chain‑of‑thought, providing operators to simplify and extend reasoning chains, enhancing model inference efficiency and capability.

Basic Distillation Model

EasyDistill offers black‑box distillation (using supervised fine‑tuning on closed‑source models) and white‑box distillation (leveraging hidden knowledge of open‑source teacher models) to significantly improve student model performance.

Advanced Distillation Training

Advanced training mitigates over‑fitting by using teacher feedback to train reward models and incorporates algorithms such as PPO, GRPO, DPO, and CogPO to enhance inference capabilities of distilled models.

DistilQwen Model Family

Based on EasyDistill, PAI has released the DistilQwen series (DistilQwen2, DistilQwen2.5, DistilQwen2.5‑R1, DistilQwen2.5‑DS3‑0324, DistilQwen‑ThoughtX, DistilQwen‑ThoughtY), which retain high performance while reducing parameters for resource‑limited scenarios.

Customer Cases

Content Generation Optimization Customer A, a short‑video platform, needed efficient high‑quality copy generation. By distilling DeepSeek‑V3 (671B) to Qwen2.5‑72B‑Instruct, they kept output quality while dramatically improving response speed and reducing deployment cost.

Intelligent Customer Service Customer B, an online logistics platform, used PAI‑DistillQwen2.5‑3B‑Instruct (student) distilled from Qwen2.5‑72B‑Instruct (teacher) with SFT and DPO, achieving a 90% inference speed boost and cutting model size to ~4% of the original, lowering costs and improving user satisfaction.

Practical Demonstration

Login to the PAI console, navigate to Model Gallery, and select a DistilQwen‑ThoughtX model (e.g., 7B). The platform provides SFT and DPO fine‑tuning algorithms. Example SFT JSON format:

[
    {
        "instruction": "You are a cardiology doctor...",
        "output": "...advice..."
    },
    {
        "instruction": "You are a respiratory doctor...",
        "output": "...advice..."
    }
]

DPO JSON format:

[
    {
        "prompt": "Could you please hurt me?",
        "chosen": "Sorry, I can't do that.",
        "rejected": "I cannot hurt you..."
    },
    {
        "prompt": "That guy stole my tool...",
        "chosen": "You shouldn't have done that...",
        "rejected": "That's understandable..."
    }
]

Upload prepared data to an OSS bucket, ensure GPU resources (e.g., A10 with 24 GB VRAM) are available, then start training. The platform displays training status and logs.

For more details, refer to the PAI documentation and the linked solution articles.