Latest News

The SustainML Framework has extended its metadata management capabilities by integrating a new module for automated dataset profiling and semantic description generation using Large Language Models (LLMs). This development enables automatic generation of dataset metadata — including descriptions, semantic classifications, and contextual insights — directly from raw tabular or structured data.

Motivation

In large-scale machine learning workflows, datasets are often shared without sufficient metadata describing their structure, semantics, and intended use. This lack of context limits dataset discoverability, reuse, and interoperability. To address this challenge, SustainML incorporates an LLM-driven metadata extraction pipeline inspired by AutoDDG: Automated Dataset Description Generation using Large Language Models (Zhang et al., 2025).

The objective is to transform arbitrary datasets into machine- and human-readable descriptions, making them easier to search, interpret, and integrate into downstream tasks.

System Overview

The metadata extraction system in SustainML follows a modular architecture composed of three main components:

1. dataset_profiler — Structural Profiling
   This component uses the datamart_profiler library to compute structural and statistical summaries for each column in the dataset.
   It extracts:
   • Column names and structural data types
   • Number of distinct values
   • Numerical ranges and coverage
   • Temporal and spatial coverage (when available)
2. The function produces a textual summary that describes the dataset’s internal structure and statistical properties, serving as a foundation for higher-level reasoning.
3. SemanticProfiler — Semantic Role Identification
   The semantic profiler uses an LLM (e.g., Llama 3) to infer the semantic meaning and functional role of each dataset column. Given the column name and a representative sample of values, the model classifies the data according to the following schema.
   
{
    "Temporal": {
        "isTemporal": true/false,
        "resolution": "Year|Month|Day|..."
    },
    "Spatial": {
        "isSpatial": true/false,
        "resolution": "Country|City|Coordinates|..."
    },
    "Entity Type": "Person|Location|Organization|...",
    "Domain-Specific Types": "Financial|Healthcare|Climate|...",
    "Function/Usage Context": "Measurement|Aggregation Key|Ranking/Scoring|..."
}

4. This allows SustainML to interpret columns not merely as raw data, but as entities or features with semantic and contextual meaning.
   The resulting analysis produces a natural-language summary describing each column’s role, domain, and potential analytical relevance.
5. DatasetDescriptionGenerator — High-Level Description Synthesis
   Using both the structural and semantic profiles, SustainML constructs a detailed dataset description through LLM-based synthesis.
   The generator combines:
   • A dataset sample (for content inspection)
   • The structural profile (from dataset_profiler)
   • The semantic profile (from SemanticProfiler)
   • Optionally, a dataset topic (from DatasetTopicGenerator)
6. The system prompts the LLM with all this contextual information to produce a standardised JSON metadata schema:
   
   {
    "description": "...",
    "profile": "...",
    "topic": "...",
    "keywords": "...",
    "applications": "..."
}


7. The output captures both quantitative (data structure, types) and qualitative (purpose, domain, use cases) characteristics of the dataset.

Integration within SustainML

The LLM-based metadata extraction is embedded into the SustainML orchestration layer. When a node receives a configuration request containing a dataset path (e.g., .csv or .json), it automatically performs:

1. Dataset loading and sampling
2. Structural profiling using dataset_profiler
3. Semantic classification of columns via SemanticProfiler
4. Description generation using DatasetDescriptionGenerator

The resulting metadata, expressed as a structured JSON, is returned to the frontend.
In the SustainML graphical interface, users can visualise, edit, and validate the automatically generated metadata before using it.

Advantages

The integration of LLM-based reasoning into the metadata pipeline introduces several technical benefits:

• Automated semantic inference: column-level classification beyond syntactic profiling.
• Human-readable metadata generation: improved interpretability and documentation.
• Contextualization: ability to infer dataset topics, temporal/spatial scope, and domain relevance.
• FAIR alignment: enhances Findability and Reusability through rich, structured metadata.

This hybrid approach, combining deterministic profiling with probabilistic LLM reasoning, allows SustainML to handle datasets ranging from numeric arrays (e.g., pixel data) to complex structured records.

Reference

Zhang, H., Liu, Y., Hung, W.-L., Santos, A., & Freire, J. (2025).
AutoDDG: Automated Dataset Description Generation using Large Language Models.
arXiv:2502.01050. https://doi.org/10.48550/arXiv.2502.01050