Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- text-generation
|
| 5 |
+
language:
|
| 6 |
+
- en
|
| 7 |
+
tags:
|
| 8 |
+
- structured-data
|
| 9 |
+
- sft
|
| 10 |
+
- synthetic
|
| 11 |
+
- json
|
| 12 |
+
- xml
|
| 13 |
+
- yaml
|
| 14 |
+
- toml
|
| 15 |
+
- csv
|
| 16 |
+
size_categories:
|
| 17 |
+
- 1K<n<10K
|
| 18 |
+
---
|
| 19 |
+
|
| 20 |
+
# 5k Mixed Hard-Structured SFT Dataset (v1)
|
| 21 |
+
|
| 22 |
+
This dataset contains **5,000 synthetic samples** designed to improve LLM performance on complex structured data conversion, extraction, and formatting tasks.
|
| 23 |
+
|
| 24 |
+
It aggregates 13 distinct conversion tasks with a specific focus on format diversity and structural complexity.
|
| 25 |
+
|
| 26 |
+
## Dataset Summary
|
| 27 |
+
|
| 28 |
+
The dataset is distributed across five major formats with the following allocation:
|
| 29 |
+
|
| 30 |
+
| Target Format | Count | Share | Task Types |
|
| 31 |
+
| :--- | :--- | :--- | :--- |
|
| 32 |
+
| **YAML** | 1,500 | 30% | `xml_to_yaml`, `text_to_yaml`, `json_to_yaml`, `csv_to_yaml`, `toml_to_yaml` |
|
| 33 |
+
| **TOML** | 1,500 | 30% | `text_to_toml` (Extraction focus) |
|
| 34 |
+
| **XML** | 1,000 | 20% | `json_to_xml`, `csv_to_xml` |
|
| 35 |
+
| **JSON** | 500 | 10% | `text_to_json`, `csv_to_json`, `xml_to_json`, `yaml_to_json`, `toml_to_json` |
|
| 36 |
+
| **CSV** | 500 | 10% | `text_to_csv`, `json_to_csv`, `xml_to_csv`, `yaml_to_csv` |
|
| 37 |
+
|
| 38 |
+
## Data Format
|
| 39 |
+
|
| 40 |
+
Each sample is in JSONL format with the following structure:
|
| 41 |
+
- `id`: A unique hash derived from the content.
|
| 42 |
+
- `category`: The high-level task category (e.g., `C_XML`, `C_TOML`).
|
| 43 |
+
- `subcategory`: Specific conversion/extraction type (e.g., `xml_to_yaml`).
|
| 44 |
+
- `task`: Action (`transform` or `extract`).
|
| 45 |
+
- `seed`: Marker for generation source (`dummy_hard` or `bench_fill`).
|
| 46 |
+
- `messages`: Conversation format (User prompt and Assistant response).
|
| 47 |
+
|
| 48 |
+
## Features
|
| 49 |
+
|
| 50 |
+
- **Strict XML Structure**: XML outputs use flattened list representations (repeating tags) rather than generic `<item>` wrappers, ensuring better semantic alignment.
|
| 51 |
+
- **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.
|
| 52 |
+
- **Deterministic Validity**: All samples are strictly validated against standard parsers (PyYAML, tomllib, ElementTree, csv).
|
| 53 |
+
|
| 54 |
+
## License
|
| 55 |
+
|
| 56 |
+
This dataset is licensed under **CC-BY-4.0** as it is a purely synthetic collection.
|
| 57 |
+
|
| 58 |
+
## Citation
|
| 59 |
+
|
| 60 |
+
If you use this dataset in your research, please cite the StructEval-T project.
|
| 61 |
+
|
| 62 |
+
---
|
| 63 |
+
*Note: This dataset is intended for supervised fine-tuning (SFT) of LLM agents to improve their structural reasoning and formatting capabilities.*
|