Datasets:

daichira
/

structured-5k-mix-sft

license: cc-by-4.0
task_categories:
  - text-generation
language:
  - en
tags:
  - structured-data
  - sft
  - synthetic
  - json
  - xml
  - yaml
  - toml
  - csv
size_categories:
  - 1K<n<10K

5k Mixed Hard-Structured SFT Dataset (v1)

This dataset contains 5,000 synthetic samples designed to improve LLM performance on complex structured data conversion, extraction, and formatting tasks.

It aggregates 13 distinct conversion tasks with a specific focus on format diversity and structural complexity.

Dataset Summary

The dataset is distributed across five major formats with the following allocation:

Target Format	Count	Share	Task Types
YAML	1,500	30%	`xml_to_yaml`, `text_to_yaml`, `json_to_yaml`, `csv_to_yaml`, `toml_to_yaml`
TOML	1,500	30%	`text_to_toml` (Extraction focus)
XML	1,000	20%	`json_to_xml`, `csv_to_xml`
JSON	500	10%	`text_to_json`, `csv_to_json`, `xml_to_json`, `yaml_to_json`, `toml_to_json`
CSV	500	10%	`text_to_csv`, `json_to_csv`, `xml_to_csv`, `yaml_to_csv`

Data Format

Each sample is in JSONL format with the following structure:

id: A unique hash derived from the content.
category: The high-level task category (e.g., C_XML, C_TOML).
subcategory: Specific conversion/extraction type (e.g., xml_to_yaml).
task: Action (transform or extract).
seed: Marker for generation source (dummy_hard or bench_fill).
messages: Conversation format (User prompt and Assistant response).

Features

Strict XML Structure: XML outputs use flattened list representations (repeating tags) rather than generic <item> wrappers, ensuring better semantic alignment.
Clean TOML/YAML Extraction: Text-to-Structured tasks output only the requested attributes as flat key-value pairs or dotted tables, avoiding unnecessary root wrappers.
Deterministic Validity: All samples are strictly validated against standard parsers (PyYAML, tomllib, ElementTree, csv).

License

This dataset is licensed under CC-BY-4.0 as it is a purely synthetic collection.

Citation

If you use this dataset in your research, please cite the StructEval-T project.

Note: This dataset is intended for supervised fine-tuning (SFT) of LLM agents to improve their structural reasoning and formatting capabilities.