DeepPrep: An LLM-Powered Agentic System for Autonomous Data Preparation

📄 Paper: This work is published at PVLDB 2026 (Volume 19, Issue 1).

DeepPrep is an LLM-powered agentic system designed for Autonomous Data Preparation (ADP). It transforms heterogeneous and noisy raw tables into analysis-ready data based on high-level natural language specifications.

Unlike traditional linear interaction methods (e.g., ReAct) that struggle with error propagation, DeepPrep introduces a Tree-based Agentic Reasoning framework. This allows the system to explore, backtrack, and revise decisions based on intermediate execution feedback.

✨ Key Features

• 🌲 Tree-Based Agentic Reasoning: Organizes pipeline construction as a search tree. The agent iteratively performs <plan>, <expand>, and <execute> actions, enabling non-local backtracking when downstream execution fails.
• 🧠 Progressive Agentic Training: A novel training framework combining:
  • Cold-Start Curriculum: Initializes the model with operator syntax and reasoning procedures.
  • Multi-turn RL with Hybrid Reward: Optimizes the policy using outcome correctness, partial progress, and process-level rewards.
• 🔄 Execution-Grounded Environment: Maintains materialized intermediate table states and provides rich runtime feedback (e.g., error traces, data samples) to guide the agent.
• 🚀 High Performance & Low Cost: DeepPrep achieves accuracy comparable to GPT-5 at 15× lower inference cost.
• 🛠 Comprehensive Operator Support: Supports 31+ operators covering cleaning, joining, aggregation, reshaping, and Python code synthesis.

📊 Performance

DeepPrep establishes state-of-the-art performance among open-source baselines and rivals proprietary models.

Datasets

We evaluate on three datasets:

Dataset	Train	Test	Pipeline Length	# Op Types
Synth-Spider (In-domain)	6,788	2,908	1-28	31
Synth-Bird (Out-of-domain)	782	1,135	2-25	31
Parrot	13,965	1,365	1-17	17

Evaluation on Synth-Spider (In-Domain) & Synth-Bird (Out-of-Domain)

Method	Backbone	Synth-Spider (Acc)	Synth-Bird (Acc)
DeepPrep (Ours)	Qwen3-14B	67.18	54.09
DeepPrep (Ours)	Qwen3-8B	65.99	53.39
ReAct	GPT-5	67.03	-
ReAct	Qwen3-14B	40.39	16.04
CodeGen	Qwen3-14B	45.47	29.48
AutoPrep	Qwen3-14B	36.75	10.41

Note: DeepPrep achieves accuracy comparable to GPT-5 at 15× lower inference cost

📁 Project Structure

• app/: Execution Environment (Sandbox) handling table states and operator execution.
• chatapp/: Interactive web interface for ADP tasks.
• _config/: Configuration files.
• src/: Core implementation of Tree-based Agentic Reasoning and RL training.
• data_synthesis/: Scripts for generating Synth-Spider and Synth-Bird datasets.

🚀 Quick Start

1. Setup

1.1 Environment Installation

conda create -n deepprep python==3.11.11
conda activate deepprep
pip install -r requirements.txt

1.2 Dataset Download

Download the synthesized datasets (Synth-Spider, Synth-Bird) from our Hugging Face repository and extract them.

1.3 Configuration

1. Base Config: Update _config/base_config.yaml:

   • openai_base_url: URL for your LLM inference server (e.g., vLLM).
   • nl2sql_data_root: Path to the downloaded datasets.

2. API Keys: Add keys to /_keys/keys.txt if using proprietary models or authenticated endpoints.

3. Model Config: Create/Edit a config file (e.g., _config/deepprep_qwen3_14b.yaml):

   agent_max_err_cnt: 5
   execute_mode: rule # or 'code' for python execution
   framework: tree_search # Enables Tree-based Agentic Reasoning
   llm_name: Qwen/Qwen2.5-14B-Instruct # or your local model path
   max_explore_turn: 5

2. Usage

Run Evaluation

To evaluate DeepPrep on the benchmark datasets:

python example/mulprocess_eval/main.py \
  --cfg deepprep_qwen3_14b \
  --benchmark synth-spider \
  --split test \
  --n 8

Interactive Demo

Launch the web UI to interact with DeepPrep:

1. Start the backend server (see App README).
2. Start the frontend (see ChatApp README).

🦁 Model Zoo

We release a comprehensive suite of trained models ranging from 0.5B to 14B parameters, based on Qwen2.5 and Qwen3.

Model	Parameters	Base Model	Link
DeepPrep-Qwen3-14B	14B	Qwen3-14B	🤗 HuggingFace
DeepPrep-Qwen3-8B	8B	Qwen3-8B	🤗 HuggingFace
DeepPrep-Qwen3-4B	4B	Qwen3-4B	🤗 HuggingFace
DeepPrep-Qwen3-1.7B	1.7B	Qwen3-1.7B	🤗 HuggingFace
DeepPrep-Qwen3-0.6B	0.6B	Qwen3-0.6B	🤗 HuggingFace
DeepPrep-Qwen2.5-14B	14B	Qwen2.5-14B	🤗 HuggingFace
DeepPrep-Qwen2.5-7B	7B	Qwen2.5-7B	🤗 HuggingFace
DeepPrep-Qwen2.5-3B	3B	Qwen2.5-3B	🤗 HuggingFace
DeepPrep-Qwen2.5-1.5B	1.5B	Qwen2.5-1.5B	🤗 HuggingFace
DeepPrep-Qwen2.5-0.5B	0.5B	Qwen2.5-0.5B	🤗 HuggingFace

🖋 Citation

If you find DeepPrep useful for your research, please cite our work:

@article{fan2026deepprep,
  title={DEEPPREP: An LLM-Powered Agentic System for Autonomous Data Preparation},
  author={Fan, Meihao and Fan, Ju and Zhang, Yuxin and Zhang, Shaolei and Du, Xiaoyong and Song, Jie and Li, Peng and Jiang, Fuxin and Zhang, Tieying and Chen, Jianjun},
  journal={PVLDB},
  volume={19},
  number={1},
  year={2026}
}

PVLDB Reference Format:

Meihao Fan, Ju Fan, Yuxin Zhang, Shaolei Zhang, Xiaoyong Du, Jie Song, Peng Li, Fuxin Jiang, Tieying Zhang, Jianjun Chen. 2026. DEEPPREP: An LLM-Powered Agentic System for Autonomous Data Preparation. PVLDB, 19(1): XXX-XXX. DOI: https://doi.org/XXX-XXX-XX

🤝 Acknowledgements

We thank the developers of Qwen, Spider, and BIRD for their open-source contributions which facilitated this research.