data_science.md

December 3, 2025 · View on GitHub

Alias for Data Science

An autonomous agent that runs your entire data science workflow.

Overview

Alias-DataScience is an autonomous, ready-to-use, intelligent assistant for real-world data science workflows. It transforms high-level analytical questions into executable plans, which can seamlessly handle data acquisition, cleaning, modeling, visualization, and narrative reporting, with minimal human intervention.

✨ Key Features

🔍 Scalable File Filtering

To handle massive data files commonly found in enterprise data lakes, Alias-DataScience combines parallelized grep operations with Retrieval-Augmented Generation (RAG) to build a low-latency, high-throughput file filtering pipeline. This preprocessing step enables accurate identification of relevant files, significantly expanding our scope and applicability.

🧠 Context-Aware Prompt Engineering

Rather than relying on generic instructions, Alias-DataScience employs three specialized prompt templates, each fine-tuned for a dominant data science workflow:

Exploratory Data Analysis (EDA): Surfaces trends, anomalies, and relationships to answer "what's happening?" and "why?"
Predictive Modeling: Automates feature engineering, model selection, and optimization.
Exact Data Computation: Delivers precise, auditable answers to quantitative queries (e.g., "What was the YoY revenue growth in Q3?").

An intelligent prompt selector routes tasks to the best template based on user intent.

📊 Handling of Messy Tabular Data

Alias-DataScience parses irregular spreadsheets (merged cells, embedded notes, multi-level headers) and converts them into structured tables. For large files, it outputs a semantic-preserving JSON representation, enabling reliable analysis of human-crafted inputs.

👁️ Multimodal Understanding of Visual Content

Image Understanding: Interprets charts, diagrams, and general images to extract numerical data, trends, and domain-specific entities
Visual QA: Answers natural-language questions about visual elements (e.g., "What was the peak value in Q3?").

📑 Automated Reporting

For EDA tasks, Alias-DataScience generates an interactive HTML report featuring:

Actionable insights backed by statistics and visuals,
Executable code snippets for transparency and reuse.

This bridges the gap between data scientists and stakeholders like business users or auditors.

📈 Benchmark Performance

Alias-DataScience achieves state-of-the-art (SOTA) across major data science agent benchmarks.

DSBench

Realistic tasks from ModelOff & Kaggle; includes multimodal inputs, multi-source data, and large-scale modeling.

Task Category	Framework	Model	Score
Data Analysis	Alias-DataScience	Qwen3-max-Preview	55.58% 🏆
	AutoGen	GPT-4	30.69%
	AutoGen	GPT-4o	34.12%
	CodeInterpreter	GPT-4	26.39%
	CodeInterpreter	GPT-4o	23.82%
Data Modeling	Alias-DataScience	Qwen3-max-Preview	49.70% 🏆
	AutoGen	GPT-4	45.52%
	AutoGen	GPT-4o	34.74%
	CodeInterpreter	GPT-4	26.14%
	CodeInterpreter	GPT-4o	16.90%

InsightBench

Open-ended comprehensive analytical tasks.

Framework	Model	Score
Alias-DataScience	Qwen3-max-Preview	43.29% 🏆
AgentPoirot	Qwen3-max-Preview	39.30%

DABench

End-to-end data analysis from real-world CSVs.

Framework	Model	Score
Alias-DataScience	Qwen3-max-Preview	95.20% 🏆
AutoGen	GPT-4	71.49%
Data Interpreter	GPT-4	73.55%
Data Interpreter	GPT-4o	94.93%

Some tables include data from published sources, used with gratitude to the original authors and cited in good faith. For accuracy, please refer to the original publications.