Datasets:

DJLougen
/

Acta

language:
  - en
license: apache-2.0
library_name: datasets
tags:
  - agentic
  - tool-use
  - function-calling
  - curated
  - high-quality
  - conversations
size_categories:
  - 1K<n<10K

Acta

Acta (Latin: "acts, deeds, records") - A premium curated sample of high-quality agentic tool-use conversations, filtered using an 8-factor quality model based on statistical correlation analysis of diversity metrics.

Overview

This public sample contains 5,000 conversations from the full 37K Acta dataset. It was created by analyzing 5 major agentic datasets (FlameFox, Hermes Reasoning, Hermes Multi-turn, Smolagents, SWE Agent) and applying evidence-based quality filters to maximize training signal density.

Key Statistics:

Train: 4,500 samples
Validation: 500 samples
Avg semantic diversity: 0.53
Avg lexical diversity: 0.47
Sources: 5 datasets harmonized

The 8-Factor Quality Model

Each sample was scored using 8 factors weighted by their correlation with training quality:

Factor	Weight	Threshold	Rationale
Lexical Diversity	0.25	>=0.35	unique_words / total_words (r=+0.967 with quality)
Semantic Diversity	0.20	>=0.40	content richness excluding stopwords (r=+0.947)
Verb Uniqueness	0.15	>=0.01	unique_verbs / total_verbs (r=+0.852)
Turn Efficiency	0.15	2-12 turns	Optimal conversation length (longer = repetitive)
Tool Pattern Novelty	0.10	-	Penalty for over-represented tool sequences
Reasoning Density	0.08	-	Thinking blocks per assistant turn
Role Balance	0.05	-	User/assistant turn ratio
Content Brevity	0.02	-	Information density (shorter = denser signal)

Source Distribution

Source	Samples	Avg Quality	Characteristics
FlameFox Agentic	~2,600	0.819	Highest diversity, balanced tools
Hermes Reasoning	~1,400	0.598	Good semantic diversity
Hermes Multi-turn	~450	0.534	Multi-turn, deduplicated
Smolagents Code	~70	0.592	Low redundancy
SWE Agent GLM	~20	0.433	Shortest, least repetitive traces

Usage

from datasets import load_dataset

# Load the curated sample
dataset = load_dataset("DJLougen/Acta")

# Access quality metrics
sample = dataset["train"][0]
print(sample["quality_score"])  # 0.0 - 1.0
print(sample["semantic_diversity"])  # 0.546
print(sample["lexical_diversity"])  # 0.400

Key Findings

Quality ≠ Quantity - 37K curated samples > 200K raw samples
Lexical diversity is the strongest quality predictor (r=+0.967)
More content ≠ better signal - SWE Agent has 5x more text but 3x lower quality
4-10 turns optimal - longer conversations become repetitive (r=-0.982)
Tool redundancy is rampant - some patterns repeat 3000+ times in raw data

Citation

@dataset{acta_2026,
  title = {Acta: A Quality-Curated Agentic Tool-Use Dataset},
  author = {Lougen, Daniel},
  year = {2026},
  url = {https://huggingface.co/datasets/DJLougen/Acta}
}

License

Apache 2.0

Acknowledgments

Source datasets:

Full 37K proprietary dataset: DJLougen/Acta-Proprietary