DeepAnalyze Documentation

ms-swift (Scalable lightWeight Infrastructure for Fine-Tuning) is a framework for training, inference, evaluation, quantization, and deployment of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs).

Key Features:

• Model Support: 500+ text models (e.g., Qwen, DeepSeek-R1, Llama) and 200+ multimodal models (e.g., Qwen2.5-VL, LLaVA).
• Training Techniques: Supports LoRA, QLoRA, GaLore, UnSloth, Liger, and more.
• Human Alignment (RLHF): Supports DPO, GRPO, PPO, KTO, CPO, SimPO, and ORPO.
• Inference Acceleration: Integrates with vLLM, SGLang, and LMDeploy.
• Quantization: Supports GPTQ, AWQ, and BNB.
• Distributed Training: Supports DDP, DeepSpeed (ZeRO2/3), FSDP, and Megatron.
• Multimodal Support: Training for VQA, Caption, OCR, and Grounding tasks.

SkyAgent is a generic agent layer designed for researchers to implement, train, and evaluate agentic tasks. It provides a modular interface that allows for:

• Implementing custom tasks.
• Using various training backends or running standalone evaluations.
• Modifying runtime implementations for specific tasks.
• Improving dispatching logic for batches of trajectories.

SkyRL is a full-stack Reinforcement Learning (RL) library designed for Large Language Models (LLMs). It consists of three primary components:

• skyagent: The agent layer used for training long-horizon, real-world agents (includes code for SkyRL-v0).
• skyrl-train: A modular and performant training framework specifically for RL.
• skyrl-gym: A collection of tool-use tasks (math, coding, search, and SQL) implemented using the Gymnasium API.

SWIFT (Scalable lightWeight Infrastructure for Fine-Tuning) is a framework for fine-tuning and deploying large language models (LLMs) and multi-modal large models.

Key Capabilities:

• Model Support: 500+ text models (e.g., Qwen, DeepSeek-R1, Llama) and 200+ multi-modal models (e.g., Qwen2.5-VL, Llava).
• Training Techniques: Supports lightweight methods like LoRA, QLoRA, GaLore, and UnSloth, as well as human alignment (RLHF) methods like DPO, GRPO, PPO, and ORPO.
• Inference Acceleration: Integrates with vLLM, SGLang, and LMDeploy.
• Quantization: Supports GPTQ, AWQ, and BNB.
• Distributed Training: Supports DDP, DeepSpeed (ZeRO2/ZeRO3), FSDP, and Megatron parallelism.

SkyRL-Gym provides several types of environments for training and evaluating agentic models. These environments range from single-domain tasks to complex multi-tool and multi-domain scenarios:

• Basic Environments: Focused on single domains such as Search, SQL, Math, or Simple Code (LCB).
• Multi-Tool Environment: The model must decide which tool to call at each step. This environment uses different tool parsing logic compared to basic environments. A common example is a combination of Search and Python Code.
• Mix Dataset Environment: Designed for multi-domain datasets where the agent must switch between different environment initializations. This requires reusing the environment and utilizing a different agent loop.
• Replicate Results: Specific configurations used to replicate known results, such as Sky-SQL-RL-7B and Search-r1 3B.

DSBench is a benchmark designed to evaluate data science agents using realistic data analysis and data modeling tasks sourced from ModelOff and Kaggle.

Input:

• Task instruction (which may include images and tables).
• Data files.

Goal:

• The agent must generate a solution that resolves the described task.

SkyRL-Gym is a library of Reinforcement Learning (RL) environments specifically designed for Large Language Models (LLMs), implemented using the Gymnasium API.

Key capabilities include:

• A standard Environment interface following the Gymnasium API.
• Ready-built environments for specialized tasks: math, code, search, and text-to-SQL.
• A reusable tool interface that allows developers to implement a tool once and use it across multiple environments.
• Support for multi-tool environments.

ms-swift is a comprehensive framework for training (CPT, SFT, RLHF), inference, evaluation, quantization, and deployment of over 500 LLMs and 200+ MLLMs.

Key capabilities include:

• Model Support: Pure text, multi-modal, sequence classification, and embedding models.
• Training Methods: Lightweight fine-tuning (LoRA, QLoRA, DoRA, etc.) and distributed training (DDP, DeepSpeed, FSDP, Megatron).
• RLHF: Supports DPO, GRPO, PPO, KTO, and more.
• Inference Acceleration: Supports PyTorch, vLLM, SGLang, and LmDeploy.
• Quantization: Supports exporting models in AWQ, GPTQ, and BNB formats.

SWIFT supports training for two main categories of Reranker models:

1. ModernBERT Reranker Models: Classification-based architecture (e.g., gte-reranker-modernbert-base).
2. Qwen3-Reranker Models: Generative-based architecture (CausalLM) available in 0.6B, 4B, and 8B sizes.