---
language:
- en
license: mit
tags:
- knowledge-graph
- rdf
- information-extraction
- uk-government-contracts
- procurement
- text-to-kg
- nlp
- trustworthy-ai
task_categories:
- text-generation
- token-classification
size_categories:
- 10K<n<100K
---

# Text-to-KG Construction Dataset (UK Government Contracts)

## Dataset Summary

This dataset contains **9,244 verified UK government procurement contracts** paired with structured **RDF knowledge graph triples**, constructed for the task of automated Text-to-KG extraction. It was developed as part of a UEL–Depixen industrial placement research project focused on building **trustworthy, hallucination-free domain-specific SLMs**.

This dataset was used to fine-tune:
- 👉 [BSVGK/phi35-mini-lora-text2kg-merged](https://huggingface.co/BSVGK/phi35-mini-lora-text2kg-merged) — Zero hallucination across 1,387 unseen contracts
- 👉 [BSVGK/phi35-mini-lora-text2kg-adapter](https://huggingface.co/BSVGK/phi35-mini-lora-text2kg-adapter) — LoRA adapter

## Dataset Details

| Property | Value |
|----------|-------|
| Domain | UK Government Procurement Contracts |
| Total Samples | 9,244 training + 1,387 test |
| Format | Contract text → RDF Triples |
| Language | English |
| Source | UK Government procurement data |
| License | MIT |

## Dataset Structure

Each sample contains:

{
  "input": "Raw UK government contract text...",
  "output": [
    {"subject": "entity_1", "predicate": "relation", "object": "entity_2"},
    {"subject": "entity_1", "predicate": "relation", "object": "entity_3"}
  ]
}

## Construction Process

1. **Data Collection** — UK government procurement contracts collected from public sources
2. **Preprocessing** — Cleaning, deduplication, and normalisation of contract text
3. **Triple Extraction** — Manual and automated RDF triple annotation
4. **Verification** — Each triple verified against source contract text
5. **Quality Control** — Dual-level hallucination check (L1: relation validity, L2: entity grounding)

## Hallucination Evaluation Framework

This dataset was evaluated using a **novel dual-level hallucination framework**:

- **L1 — Relation Validity:** All relations verified against a predefined ontology
- **L2 — Entity Grounding:** All entities grounded in the source contract text

This ensured **zero hallucination** in the fine-tuned Phi-3.5 model across 1,387 unseen test contracts.

## Models Trained on This Dataset

| Model | F1 | BERTScore | Hallucination Rate |
|-------|----|-----------|--------------------|
| Phi-3.5 Mini Instruct (LoRA) | **0.9954** | **0.9997** | **0.00%** |
| Gemma 2 2B IT (QLoRA) | competitive | competitive | higher |

## Intended Use

- Training SLMs for knowledge graph construction
- Research in trustworthy and hallucination-free NLP
- Information extraction from legal and procurement documents
- RDF triple generation for semantic web applications

## Out of Scope

- Non-English contracts
- Contracts outside UK government procurement domain
- General purpose NLP tasks

## Citation

@misc{bubathula2026texttokg_dataset,
  author = {Sai Venkata Gopala Krishna Bubathula},
  title = {Text-to-KG Construction Dataset: UK Government Procurement Contracts for RDF Triple Extraction},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/BSVGK/Text_to_KG_Construction_Dataset},
  institution = {University of East London & Depixen}
}

## Developer

**Sai Venkata Gopala Krishna Bubathula**
- 🎓 MSc Big Data Technologies, University of East London
- 🏢 AI Engineer — UEL–Depixen Industrial Placement
- 🔗 [GitHub](https://github.com/BSVGK1919)
- 🔗 [LinkedIn](https://www.linkedin.com/in/sai-venkata-gopala-krishna-bubathula-a05a26283/)
- 🔗 [HuggingFace](https://huggingface.co/BSVGK)