EasyGAD: A One-stop Implementation Solution for Black-box Distillation of Large Language Models

Basic Information

• Author/Maintainer: Dr-Corgi
• Source: GitHub (https://github.com/Dr-Corgi/EasyGAD)
• Release Time: 2026-06-11
• Paper: Black-Box On-Policy Distillation of Large Language Models (arXiv:2511.10643)

Core Overview

EasyGAD is a GAD (Generative Adversarial Distillation)-based framework for LLM black-box distillation. It enables student models to learn the teacher's style and capabilities via only the teacher's text outputs (no internal parameters/logits access). Its key value lies in one-stop implementation—integrating all dependencies into a single repo for out-of-the-box use.

Traditional knowledge distillation requires access to the teacher model's internal parameters or logits, which is impossible for closed commercial models like GPT-4 or Claude.

Black-box distillation solves this by training students using only the teacher's text outputs. This allows enterprises and researchers to leverage top commercial models' capabilities to train lightweight, cost-effective local models.

GAD Core Idea

GAD (Generative Adversarial Distillation) introduces adversarial training into distillation:

• Student model as Generator: generates responses.
• Discriminator: distinguishes student vs teacher outputs.
• Minimax game: student aims to mimic teacher, discriminator improves distinction.
• Advantage: Discriminator acts as an on-policy reward model, providing stable adaptive feedback (better than SeqKD).

EasyGAD Architecture

• VeRL Core Library: Includes unified trainer (ray_trainer.py) and algorithms (core_algos.py: PPO, SFT, GRPO).
• Actor & Critic: Actor generates responses; Critic/Discriminator evaluates quality (co-optimized via adversarial training).
• DeepScaler Tools: Global config, evaluation prompts, reward functions.

Training Stages

1. SeqKD: Optional baseline (supervised fine-tuning on teacher data).
2. Warmup: Required (init discriminator, SFT student on teacher data).
3. GAD: Core stage (PPO + advantage estimation, discriminator provides rewards).
4. Eval: No parameter updates (model assessment).

Environment Preparation

• Docker Support: Recommended image czwin32768/verl2:v0.2.0-vllm085 (pre-installed Python3.10.12, PyTorch2.6.0, vLLM0.8.5).
• Data: LMSYS-Chat dataset, converted to Parquet via provided scripts (for efficient large-scale processing).

Training Options

• One-click: run_gad_full.sh (Warmup + GAD core stages) or run_gad_direct.sh (start from pre-trained Actor/Critic).
• Staged: run_seqkd.sh (baseline), run_warmup.sh (preheat), run_gad.sh (adversarial training).

Key Results

GAD-trained Qwen2.5-14B-Instruct student model performs comparably to teacher model GPT-5-Chat (per LMSYS-Chat auto evaluation).

Evaluation Note

ROUGE-L score of GAD may be lower than SeqKD, but ROUGE-L measures n-gram overlap (not deep semantic quality). It should only be used as a training diagnostic indicator, not a final quality metric.

Technical Value

• Lowers black-box distillation threshold for resource-limited researchers/enterprises.
• Proves adversarial training's effectiveness in text generation/distillation (extending from image domain).

Application Scenarios

• Enterprises can use public APIs of top models to get high-quality data, then train cost-effective private models via GAD.

EasyGAD is a well-designed, engineering-perfect LLM black-box distillation framework. It encapsulates complex adversarial distillation into easy-to-use tools (Docker support, one-click scripts) to reduce usage barriers.

For AI practitioners focused on model compression, knowledge distillation, or efficient deployment, EasyGAD is a valuable tool and an important engineering contribution to the adversarial distillation field.