Avanzado

Ajuste fino

El Ajuste Fino (Fine Tuning) le permite adaptar los modelos de NovaAI a su propio dominio, estilo de escritura o requisitos comerciales mediante el entrenamiento con conjuntos de datos personalizados. Esto ayuda a mejorar la coherencia de las respuestas, la precisión y el rendimiento en tareas específicas.

Overview

While base models perform well across a wide range of tasks, fine-tuning enables them to learn patterns specific to your application. Common use cases include customer support, legal assistants, medical documentation, and brand-specific content generation.

Fine-tune a model only when prompt engineering and system instructions are no longer sufficient. For many applications, carefully designed prompts provide excellent results without the additional cost of training.

Fine-tune a model only when prompt engineering and system instructions are no longer sufficient. For many applications, carefully designed prompts provide excellent results without the additional cost of training.

Fine-Tuning Workflow

Rendering diagram…

Project Structure

Rendering diagram…

Keeping training and validation datasets organized makes experiments easier to reproduce and maintain.

Training Dataset Guidelines

Requirement

Recommendation

Format

JSONL

Example Quality

High-quality, accurate responses

Dataset Size

Hundreds to thousands of examples

Consistency

Use the same writing style throughout

Validation Set

Keep separate from training data

Example Training Record

{
"messages": [
{
"role": "user",
"content": "Write a professional welcome email."
},
{
"role": "assistant",
"content": "Welcome! We're excited to have you join our platform..."
}
]
}

Each training example should represent the ideal response you expect from the model.

Before You Train

  • Clean and validate your dataset.

  • Remove duplicate or conflicting examples.

  • Use consistent formatting across all records.

  • Test prompts with the base model first.

  • Keep training and validation data separate.

Poor-quality or biased training data will produce poor-quality model outputs. Always review datasets carefully before starting a training job, as mistakes can affect every response generated by the fine-tuned model.

Poor-quality or biased training data will produce poor-quality model outputs. Always review datasets carefully before starting a training job, as mistakes can affect every response generated by the fine-tuned model.

Training Job Status

Status

Description

Queued

Waiting to begin training.

Running

Model is currently training.

Completed

Training finished successfully.

Failed

Training stopped due to an error.

Cancelled

Training was manually cancelled.

Best Practices

  • Start with a small dataset before scaling.

  • Continuously evaluate model performance.

  • Retrain only when your data changes significantly.

  • Version your datasets and trained models.

  • Monitor production results after deployment.


Next Step

Once your model has been trained and evaluated, continue to Batch Requests to learn how to efficiently process large volumes of requests in a single operation.

Documentación relacionada

¿Te resultó útil?

¿Te resultó útil?

Create a free website with Framer, the website builder loved by startups, designers and agencies.