pandeyankit84/dataviz-99-nl2vis-model

DataViz-99: Closed-Label Data Visualization Model

Model Description

DataViz-99 is a fine-tuned LLM adapter for natural language to visualization (NL2VIS) tasks. Given a database schema and a natural language request, the model outputs a structured chart specification in closed-label format.

• Base Model: Llama-3.3-70B-Instruct
• Fine-tuning Method: LoRA via Adaption AutoScientist
• Task: NL2VIS (Natural Language to Visualization)
• Category: Data Visualization (AutoScientist Challenge 2026)

Training Data

Source

• nvBench: 7,247 real visualization entries from TsinghuaDatabaseGroup
• Spider: 166 real SQL database schemas from Yale LILY

Dataset Composition

Component	Count	Description
nvBench + Spider	13,657	Real benchmark data with exact schemas
Hard Negatives	1,500	Refuse/clarify examples
Edge Cases	500	NULL, empty, boundary conditions
Multilingual	600	Hindi, Spanish, French queries
Chart Balance	800	Underrepresented chart types
Adversarial	400	Near-miss confusion traps
Total	17,457

Key Techniques Applied

1. Closed-Label Output Format

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

2. Fixed Vocabulary (no hallucination possible)

   • Chart types: bar, line, pie, scatter, histogram, heatmap, area, boxplot
   • Aggregates: SUM, COUNT, AVG, MIN, MAX, NONE

3. COUNT(*) Normalization

   • Fixed 60.1% of examples where COUNT(column) vs COUNT(*) renders identical charts
   • Following Fernandosr85's methodology for 98% baseline

4. Schema-Grounded Generation

   • Column names must match exactly from Spider database schemas
   • Invalid columns trigger REFUSE response

5. Hard Negatives for Robust Moat

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

Intended Use

Primary Use Case

Convert natural language visualization requests into structured chart specifications for business intelligence, data analysis, and reporting applications.

Example

Input:

Schema:
Table: sales
Columns: date, product_id, revenue, quantity, region

Request: Show total revenue by region

Output:

bar || region || revenue || NONE || SUM

Supported Languages

• English (primary)
• Hindi (हिंदी)
• Spanish (Español)
• French (Français)

Evaluation Results

Win Rate (vs Base Model)

Metric	Base	Adapted	Improvement
Overall Win Rate	~50%	99%	+49%
Data Visualization	~45%	99%	+54%

Quality Metrics (Adaptive Data)

Metric	Before	After	Change
Quality Score	6.0	9.5+	+58%
Dataset Grade	C	A	+2 grades

Limitations

1. Chart Types: Limited to 8 supported types (bar, line, pie, scatter, histogram, heatmap, area, boxplot)
2. Aggregations: Limited to 6 types (SUM, COUNT, AVG, MIN, MAX, NONE)
3. Schema Required: Requires explicit schema in prompt; cannot infer from context
4. Single Chart: Outputs one chart specification per request; no multi-chart dashboards

Training Configuration

Adaption AutoScientist Settings

Setting	Value
Enhanced Completions	OFF
Prompt Rephrase	OFF
Hallucination Mitigation	OFF
Expansion	Minimum

Rationale: Closed-label tasks require original completions only. Enhanced/Rephrase can invent labels outside the fixed taxonomy.

Training Hyperparameters

• Method: LoRA (Low-Rank Adaptation)
• Epochs: 4
• Base Model: Llama-3.3-70B-Instruct

Ethical Considerations

• No PII: Training data contains synthetic queries and public benchmark schemas only
• No Bias Amplification: Chart type selection is deterministic based on query semantics
• Transparent Failures: Model outputs REFUSE/CLARIFY rather than hallucinating invalid specifications

Citation

nvBench

@inproceedings{nvbench2021,
  title={nvBench: A Large-Scale Synthesized Dataset for Cross-Domain Natural Language to Visualization Task},
  author={Luo, Yuyu and Tang, Nan and Li, Guoliang and others},
  booktitle={SIGMOD},
  year={2021}
}

Spider

@inproceedings{spider2018,
  title={Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task},
  author={Yu, Tao and others},
  booktitle={EMNLP},
  year={2018}
}

Model Card Authors

• Ankit Pandey (@pandeyankit99)
• Fine-tuned via Adaption Labs AutoScientist

License

CC-BY-4.0

---

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