> ## Documentation Index
> Fetch the complete documentation index at: https://docs.greenflash.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Data Curation

> Transform real user conversations into high-quality training datasets powered by deep analysis.

Greenflash helps you build training datasets from real production conversations, filtered and organized by comprehensive quality analysis.

## Intelligent data selection

Greenflash's analysis engine identifies which conversations make good training examples and which demonstrate patterns to avoid.

### Quality-based curation

Export conversations filtered by:

* Quality scores (high for positive training, low for negative examples)
* User ratings and satisfaction metrics
* Edge cases and unusual patterns
* Specific topics or use cases

Every export includes:

* Full conversation transcript
* Quality metrics and analysis scores
* User ratings and feedback
* Topic classification and keywords
* Sentiment progression
* Safety issue flags

### Custom filtering

Build datasets that match your exact training needs:

* Filter by sentiment trajectory (improving vs. declining)
* Select specific error types (hallucinations, refusals, confusion)
* Choose conversation lengths and complexity levels
* Include only rated conversations with explicit feedback
* Combine multiple criteria for precise curation

## Training data that matters

### Positive examples

Greenflash identifies what makes conversations successful:

* High user satisfaction scores
* Positive sentiment throughout
* Successful task completion
* Efficient problem resolution
* No safety or quality issues

### Negative examples

Learn from what goes wrong:

* Hallucination occurrences with context
* Sentiment drops with triggering messages
* Failed task attempts with breakdown points
* Safety violations with specific patterns
* User frustration signals

### Balanced datasets

Greenflash helps create representative training sets:

* Topic coverage across categories
* Varied conversation complexity
* Balance of positive and negative examples
* Representative user interaction patterns

## Export capabilities

Each exported conversation includes:

* Complete message transcripts with metadata
* Quality scores and analysis results
* Sentiment, topic, and safety annotations
* Tool calls and system prompts
* User ratings and feedback

Available formats:

* CSV for analysis and filtering
* JSON for model training pipelines
* Custom schemas for your requirements

## Workflow

1. Deploy your AI with Greenflash monitoring
2. Accumulate and analyze real conversations
3. Export curated datasets based on quality metrics
4. Fine-tune models with production data
5. Measure improvement and iterate

## Privacy and compliance

* **Data ownership**: Your conversations remain yours
* **Filtering controls**: Exclude sensitive information
* **Anonymization**: Remove PII before export
* **Audit trails**: Track what data was exported when
* **Access controls**: Limit who can export data

## Common use cases

* **Fine-tuning**: Select high-quality conversations for model training
* **Evaluation**: Build test sets from real edge cases
* **Error analysis**: Export and study failure patterns
* **RLHF**: Curate examples with clear quality signals