# How to Estimate Training Time in FPT AI Studio

## Objectives

* How to **predict fine-tuning time** before launching a job.
* Understand how to check **ETA** during training.
* Plan GPU resource allocation more effectively.

## Overview

Assume fixed GPUs. Fine-tuning time depends on several key factors:

| Factor              | Description                   |
| ------------------- | ----------------------------- |
| **Model size**      | Number of parameters          |
| **Dataset size**    | Number of training samples    |
| **Sequence length** | Max token length per sample   |
| **Batch size**      | Samples processed per step    |
| **Epochs**          | Number of full dataset passes |

## Method 1: Empirical Estimation (Based on Real Benchmark)

### Description

This method estimates total training time by running a **small benchmark job** on a subset of the data and **scaling up the results**. It’s simple, practical, and highly accurate for your specific hardware setup.

### Steps to perform

1. **Run a mini benchmark**
   * Use **100-1000** samples from your dataset.
   * Train for **1 epoch** with your target hyperparameters.
   * Record the **training time** for this mini-run (e.g., total time for 1 epoch).
2. **Estimate total training time** Use a simple proportional scaling formula:

   ```
   Example:
   Training time for 1,000 samples (1 epoch) = 5 minutes
   Full dataset size = 20,000 samples
   Planned epochs = 3

   scaling_factor = dataset_size / sample_dataset_size
   scaling_factor = 20,000 / 1,000 = 20

   total_time = scaling_factor * epoch_time * epochs
   total_time = 20 * 5 minutes * 3 = 300 minutes = 5 hours
   ```

## Method 2: Online Estimation (Dynamic ETA)

### Description

Once training starts, FPT AI Studio can **dynamically estimate the remaining time (ETA)** based on the **actual average step time** during runtime.

### How it works in FPT AI Studio

* The backend automatically logs **the training time for each logging interval** (defined by your `logging_steps` setting).
* These logs are stored in the **Logs → training-**\* section of the Studio interface.
* You can **view ETA updates directly in the dashboard** or **download the full log file** for detailed analysis.

![log](/files/11101b5ef6a908936938ae5cdd86659fc4554968)

## Comparison of Both Methods

| Criteria                | Empirical Estimation | Online Estimation           |
| ----------------------- | -------------------- | --------------------------- |
| **When to use**         | Before training      | During training             |
| **Requires trial run?** | Yes                  | No (uses live data)         |
| **Main purpose**        | GPU/time planning    | Real-time progress tracking |

## Best Practices

* Start with **Empirical Estimation** for a rough pre-training plan.
* Use **Dynamic ETA** to **monitor progress and verify your estimate**.


---

# Agent Instructions: Querying This Documentation

If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter:

```
GET https://ai-docs.fptcloud.com/fpt-ai-studio/faq/how-to-estimate-training-time-in-fpt-ai-studio.md?ask=<question>
```

The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.