pandeyankit84/dataviz-99-nl2vis-dataset

DataViz-99: NL2VIS Dataset Card

Dataset Description

DataViz-99 is a closed-label instruction-tuning dataset for natural language to visualization (NL2VIS) tasks. Built from real benchmark data (nvBench + Spider), it teaches models to convert natural language queries into structured chart specifications.

Dataset Summary

Metric	Value
Total Examples	17,457
Format	JSONL (prompt/completion pairs)
Languages	English, Hindi, Spanish, French
Databases	166 real SQL schemas
License	CC-BY-4.0

Dataset Sources

Primary Sources

1. nvBench (TsinghuaDatabaseGroup)

   • 7,247 visualization entries
   • 25,750 (NL, VIS) pairs
   • 105 domains (sports, healthcare, finance, etc.)

2. Spider (Yale LILY)

   • 166 database schemas
   • Exact column names and types
   • Real-world SQL databases

Dataset Structure

Data Fields

{
  "prompt": "You are a data visualization assistant...\n\nSchema:\nTable: sales\nColumns: date, revenue, region\n\nRequest: Show revenue by region\n\nChart spec:",
  "completion": "bar || region || revenue || NONE || SUM"
}

Output Format (Closed-Label)

CHART_TYPE || X_COLUMN || Y_COLUMN || GROUP_COLUMN || AGGREGATE

Valid Values

Field	Valid Options
CHART_TYPE	bar, line, pie, scatter, histogram, heatmap, area, boxplot
X_COLUMN	Column name from schema or NONE
Y_COLUMN	Column name from schema or NONE
GROUP_COLUMN	Column name from schema or NONE
AGGREGATE	SUM, COUNT, AVG, MIN, MAX, NONE

Special Outputs

REFUSE: <reason>     # When request cannot be fulfilled
CLARIFY: <question>  # When more information needed

Dataset Composition

Category	Count	Percentage
Real nvBench + Spider	13,657	78.2%
Hard Negatives	1,500	8.6%
Chart Balance	800	4.6%
Multilingual	600	3.4%
Edge Cases	500	2.9%
Adversarial	400	2.3%

By Chart Type

Chart	Count
bar	~11,000
line	~2,000
pie	~1,500
scatter	~1,200
histogram	~800
heatmap	~500
boxplot	~300
area	~200

By Language

Language	Count
English	~16,200
Hindi	~400
Spanish	~400
French	~400

Data Collection Process

Step 1: nvBench Processing

• Downloaded nvBench.json from TsinghuaDatabaseGroup
• Extracted 7,247 visualization entries
• Mapped chart types to closed vocabulary

Step 2: Spider Schema Resolution

• Downloaded Spider database schemas (tables.json)
• Linked nvBench entries to exact Spider column names
• Validated column existence in schemas

Step 3: COUNT(*) Normalization

• Identified 60.1% of examples with COUNT(column) vs COUNT(*) ambiguity
• Normalized to consistent format (y=NONE, aggregate=COUNT)
• Following methodology from Fernandosr85's 98% submission

Step 4: Hard Negative Generation

• Non-existent column requests → REFUSE
• Invalid chart type requests → REFUSE
• Ambiguous requests → CLARIFY
• Off-topic requests → REFUSE

Step 5: Data Augmentation

• Edge cases (NULL, empty, boundary)
• Multilingual (Hindi, Spanish, French)
• Chart balancing (underrepresented types)
• Adversarial (near-miss traps)

Preprocessing

Cleaning Steps

1. Removed invalid chart types (kept only 8 supported)
2. Normalized column names to Spider schemas
3. Dropped examples with missing required fields
4. Deduplicated near-identical entries

Quality Filters

• Minimum query length: 10 characters
• Maximum 2 NL variations per nvBench entry
• Schema validation against Spider

Considerations for Using the Data

Intended Use

• Fine-tuning LLMs for NL2VIS tasks
• Business intelligence chart generation
• Data visualization assistants

Limitations

• English-centric (85% of examples)
• 8 chart types only (no advanced visualizations)
• Requires explicit schema in prompt

Biases

• Overrepresentation of bar charts (nvBench artifact)
• Limited to tabular data (no time-series specific handling)
• Business/academic domains predominant

Additional Information

Dataset Curators

• Ankit Pandey (@pandeyankit99)

Licensing

CC-BY-4.0 (following nvBench and Spider licenses)

Citation

@dataset{dataviz99,
  author = {Pandey, Ankit},
  title = {DataViz-99: NL2VIS Dataset for AutoScientist Challenge},
  year = {2026},
  publisher = {Kaggle/HuggingFace},
  note = {Built from nvBench and Spider benchmarks}
}

Links

• Kaggle Dataset: https://www.kaggle.com/datasets/pandeyankit99/dataviz-99-final-boosted
• HuggingFace Dataset: [To be uploaded]
• nvBench: https://github.com/TsinghuaDatabaseGroup/nvBench
• Spider: https://yale-lily.github.io/spider

---

AutoScientist Challenge 2026 - Data Visualization Category Powered by Adaptive Data + AutoScientist by Adaption Labs