Acta-Synthetic

High-quality synthetic agentic tool-use conversations generated with 8-factor quality control.

Overview

This dataset contains 5,000 synthetic agentic tool-use conversations generated using a template-based synthesis pipeline with strict quality controls.

Key Statistics:

• Train: 4,500 samples
• Validation: 500 samples
• Avg quality score: 0.824
• Avg semantic diversity: 0.914
• Avg lexical diversity: 0.819
• Pass rate: 96.2%

Generation Method

Conversations were generated using the 8-Factor Quality Model:

Factor	Weight	Rationale
Lexical Diversity	0.25	unique_words / total_words (r=+0.967 with quality)
Semantic Diversity	0.20	content richness (r=+0.947)
Verb Uniqueness	0.15	unique_verbs / total_verbs (r=+0.852)
Turn Efficiency	0.15	Optimal 4-10 turns
Tool Pattern Novelty	0.10	Penalty for over-represented patterns
Reasoning Density	0.08	Thinking blocks per turn
Role Balance	0.05	User/assistant ratio
Content Brevity	0.02	Information density

Quality threshold: 0.75 (only samples scoring above this were kept)

Domains

• Database queries: 15.3%
• Math problems: 14.8%
• API integration: 12.7%
• Coding tasks: 12.3%
• File processing: 11.6%
• Research: 11.1%
• Multi-step research: 11.1%
• Data analysis: 11.1%

Tools Available

• web_search - Web search capability
• wikipedia_search - Wikipedia knowledge retrieval
• python_interpreter - Code execution
• code_analyzer - Code analysis and review
• read_file - File reading
• write_file - File writing
• api_request - HTTP API calls
• database_query - SQL database queries
• calculator - Mathematical calculations
• data_transform - Data transformation

Comparison with Real Data

Validating against DJLougen/Acta (real agentic data):

Metric	Real (Acta)	Synthetic	Difference
Quality Score	0.788 ± 0.052	0.826 ± 0.038	+0.038
Acceptance Rate	-	57.2%	-
Unique Patterns	-	234	-

Synthetic data achieves higher quality scores than real curated data due to:

1. Strict 8-factor filtering during generation
2. Pattern deduplication (max 50 samples per tool pattern)
3. Optimized conversation structure (4-10 turns)
4. Controlled diversity in vocabulary and reasoning

Usage

from datasets import load_dataset

# Load synthetic dataset
dataset = load_dataset("DJLougen/Acta-Synthetic")

# Access quality metrics
sample = dataset["train"][0]
print(sample["quality_score"])  # 0.824
print(sample["semantic_diversity"])  # 0.914
print(sample["lexical_diversity"])  # 0.819
print(sample["tools_used"])  # ["python_interpreter", "calculator"]

Mixing with Real Data

from datasets import load_dataset
import random

# Load both datasets
real = load_dataset("DJLougen/Acta")["train"]
synthetic = load_dataset("DJLougen/Acta-Synthetic")["train"]

# Mix with 30% synthetic ratio
real_samples = [real[i] for i in range(min(3500, len(real)))]
synthetic_samples = [synthetic[i] for i in range(min(1500, len(synthetic)))]

mixed = real_samples + synthetic_samples
random.shuffle(mixed)

print(f"Mixed dataset: {len(mixed)} samples")

Generation Pipeline

The generation pipeline is available in the Acta repository.

Key components:

• SyntheticConversationGenerator - Template-based generation
• LLMSyntheticGenerator - LLM-enhanced generation (optional)
• AgenticDataQualityModel - 8-factor quality scoring
• SyntheticDataPipeline - End-to-end workflow

Quality Validation

All samples pass strict quality thresholds:

• Lexical diversity ≥ 0.35
• Semantic diversity ≥ 0.40
• Verb uniqueness ≥ 0.01
• Turn count: 2-12

Target quality score: ≥ 0.75

Citation

@dataset{acta_synthetic_2026,
  title = {Acta-Synthetic: High-Quality Synthetic Agentic Tool-Use Dataset},
  author = {Lougen, Daniel},
  year = {2026},
  url = {https://huggingface.co/datasets/DJLougen/Acta-Synthetic}
}

License

Apache 2.0

Related Datasets

• DJLougen/Acta - Public sample (5K real samples)
• DJLougen/Acta-Proprietary - Full 37K real samples (private)
• DJLougen/ornstein-proprietary-100k - 82K samples with 8-factor metrics