golden-dataset-curation

name	golden-dataset-curation
description	Use when creating or improving golden datasets for AI evaluation. Defines quality criteria, curation workflows, and multi-agent analysis patterns for test data.
context	fork
agent	data-pipeline-engineer
version	1.0.0
author	SkillForge AI Agent Hub
tags	["golden-dataset","curation","quality","multi-agent","langfuse",2025]
user-invocable	false

Golden Dataset Curation

Curate high-quality documents for the golden dataset with multi-agent validation

Overview

This skill provides patterns and workflows for adding new documents to the golden dataset with thorough quality analysis. It complements golden-dataset-management which handles backup/restore.

When to use this skill:

• Adding new documents to the golden dataset
• Classifying content types and difficulty levels
• Generating test queries for new documents
• Running multi-agent quality analysis

---

Content Types

Type	Description	Quality Focus
article	Technical articles, blog posts	Depth, accuracy, actionability
tutorial	Step-by-step guides	Completeness, clarity, code quality
research_paper	Academic papers, whitepapers	Rigor, citations, methodology
documentation	API docs, reference materials	Accuracy, completeness, examples
video_transcript	Transcribed video content	Structure, coherence, key points
code_repository	README, code analysis	Code quality, documentation

---

Difficulty Levels

Level	Semantic Complexity	Expected Score	Characteristics
trivial	Direct keyword match	>0.85	Technical terms, exact phrases
easy	Common synonyms	>0.70	Well-known concepts, slight variations
medium	Paraphrased intent	>0.55	Conceptual queries, multi-topic
hard	Multi-hop reasoning	>0.40	Cross-domain, comparative analysis
adversarial	Edge cases	Graceful degradation	Robustness tests, off-domain

---

Quality Dimensions

Dimension	Weight	Perfect	Acceptable	Failing
Accuracy	0.25	0.95-1.0	0.70-0.94	<0.70
Coherence	0.20	0.90-1.0	0.60-0.89	<0.60
Depth	0.25	0.90-1.0	0.55-0.89	<0.55
Relevance	0.30	0.95-1.0	0.70-0.94	<0.70

Evaluation focuses:

• Accuracy: Technical correctness, code validity, up-to-date info
• Coherence: Logical structure, clear flow, consistent terminology
• Depth: Comprehensive coverage, edge cases, appropriate detail
• Relevance: Alignment with AI/ML, backend, frontend, DevOps domains

---

Multi-Agent Pipeline

INPUT: URL/Content
        |
        v
+------------------+
|   FETCH AGENT    |  Extract structure, detect type
+--------+---------+
         |
         v
+-----------------------------------------------+
|  PARALLEL ANALYSIS AGENTS                      |
|  Quality | Difficulty | Domain  | Query Gen   |
+-----------------------------------------------+
         |
         v
+------------------+
| CONSENSUS        |  Weighted score + confidence
| AGGREGATOR       |  -> include/review/exclude
+--------+---------+
         |
         v
+------------------+
|  USER APPROVAL   |  Show scores, confirm
+--------+---------+
         |
         v
OUTPUT: Curated document entry

Decision Thresholds

Quality Score	Confidence	Decision
>= 0.75	>= 0.70	include
>= 0.55	any	review
< 0.55	any	exclude

---

Quality Thresholds

# Recommended thresholds for golden dataset inclusion
minimum_quality_score: 0.70
minimum_confidence: 0.65
required_tags: 2          # At least 2 domain tags
required_queries: 3       # At least 3 test queries

---

Coverage Balance Guidelines

Maintain balanced coverage across:

• Content types: Don't over-index on articles
• Difficulty levels: Need trivial AND hard queries
• Domains: Spread across AI/ML, backend, frontend, etc.

Duplicate Prevention Checklist

Before adding:

1. Check URL against existing source_url_map.json
2. Run semantic similarity against existing document embeddings
3. Warn if >80% similar to existing document

Provenance Tracking

Always record:

• Source URL (canonical)
• Curation date
• Agent scores (for audit trail)
• Langfuse trace ID

---

Langfuse Integration

Trace Structure

trace = langfuse.trace(
    name="golden-dataset-curation",
    metadata={"source_url": url, "document_id": doc_id}
)

# Log individual dimension scores
trace.score(name="accuracy", value=0.85)
trace.score(name="coherence", value=0.90)
trace.score(name="depth", value=0.78)
trace.score(name="relevance", value=0.92)

# Final aggregated score
trace.score(name="quality_total", value=0.87)
trace.event(name="curation_decision", metadata={"decision": "include"})

Managed Prompts

Prompt Name	Purpose
golden-content-classifier	Classify content_type
golden-difficulty-classifier	Assign difficulty
golden-domain-tagger	Extract tags
golden-query-generator	Generate test queries

---

References

For detailed implementation patterns, see:

• references/selection-criteria.md - Content type classification, difficulty stratification, quality evaluation dimensions, and best practices
• references/annotation-patterns.md - Multi-agent pipeline architecture, agent specifications, consensus aggregation logic, and Langfuse integration

---

Related Skills

• golden-dataset-management - Backup/restore operations
• golden-dataset-validation - Validation rules and checks
• langfuse-observability - Tracing patterns
• pgvector-search - Duplicate detection

---

Version: 1.0.0 (December 2025) Issue: #599

Capability Details

content-classification

Keywords: content type, classification, document type, golden dataset Solves:

• Classify document content types for golden dataset
• Categorize entries by domain and purpose
• Identify content requiring special handling

difficulty-stratification

Keywords: difficulty, stratification, complexity level, challenge rating Solves:

• Assign difficulty levels to golden dataset entries
• Ensure balanced difficulty distribution
• Identify edge cases and challenging examples

quality-evaluation

Keywords: quality, evaluation, quality dimensions, quality criteria Solves:

• Evaluate entry quality against defined criteria
• Score entries on multiple quality dimensions
• Identify entries needing improvement

multi-agent-analysis

Keywords: multi-agent, parallel analysis, consensus, agent evaluation Solves:

• Run parallel agent evaluations on entries
• Aggregate consensus from multiple analysts
• Resolve disagreements in classifications