Skip to content

Update Fine-tune with LoRA.md #11

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Thaonp114 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Update Fine-tune with LoRA.md #11

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
46 changes: 17 additions & 29 deletions Fine-tune with LoRA.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,37 +1,28 @@
# Supervised Fine-Tuning for Text (SFT)
This guide focuses on using supervised fine-tuning (SFT) to fine-tune and deploy a model with on-demand and serverless hosting.

This guide focuses on using **supervised fine-tuning** (SFT) to fine-tune and deploy a model with **on-demand** and **serverless hosting**.

---

## 📌 List of Supported Models

We currently support fine-tuning models with the following architectures:
## 1. List of Supported Models

We currently support fine-tuning models with the following architectures:
| Model (Resource) | Suggested Learning Rate (small → med) | Suggested Epochs |
|-----------------------------------------------|---------------------------------------------|------------------------------|
| **Qwen-3 / Qwen3-4B-Instruct** (template 1 GPU) | small: `1e-5 → 5e-5`<br>med: `5e-5 → 1e-4` | small: 1–3<br>med: 3–5 |
| **google/gemma-3-27b-it** (template 2 GPUs) | small: `1e-5 → 5e-5`<br>med: `5e-5 → 1e-4` | 3 (start) |
| **meta-llama/Llama-3.3-70B** (template 4 GPUs) | small: `1e-5 → 2e-5`<br>med: `2e-5 → 1e-4` | 3 (start) |

---

## 📂 Dataset Format
## 2. Dataset Format

| Dataset Type | Link to Sample | Note |
|----------------|--------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------|
| **Alpaca** | [Alpaca Sample](https://github.com/fpt-corp/ai-studio-samples/tree/main/sample-datasets/alpaca) | Dễ chuẩn bị, phù hợp với các use case instruction tuning cơ bản (summarization, QA, rewriting). |
| **ShareGPT** | [ShareGPT Sample](https://github.com/fpt-corp/ai-studio-samples/tree/main/sample-datasets/sharegpt) | Phù hợp khi người dùng muốn fine-tune chatbot có khả năng hội thoại nhiều vòng. |
| **ShareGPT_Image** | [ShareGPT_Image Sample](https://github.com/fpt-corp/ai-studio-samples/tree/main/sample-datasets/sharegpt-image) | Dành cho nhóm user nâng cao, làm về AI đa phương tiện. |

---

## 💡 What is Training Data?
| **Alpaca** | [Alpaca Sample](https://github.com/fpt-corp/ai-studio-samples/tree/main/sample-datasets/alpaca) | Easy to prepare, suitable for basic instruction tuning use cases (summarization, QA, rewriting). |
| **ShareGPT** | [ShareGPT Sample](https://github.com/fpt-corp/ai-studio-samples/tree/main/sample-datasets/sharegpt) | Suitable when users want to fine-tune a chatbot with multi-turn conversation capability. |
| **ShareGPT_Image** | [ShareGPT_Image Sample](https://github.com/fpt-corp/ai-studio-samples/tree/main/sample-datasets/sharegpt-image) | Designed for advanced users working on multimedia AI. |

In the fine-tuning process, **data** refers to a curated set of example inputs and outputs used to retrain a pre-trained AI model.
This data teaches the model to adapt its behavior to suit your **specific domain, task, or tone of voice**.

## **1. Alpaca**
### **2.1 Alpaca**

**Alpaca** uses a very simple structure to fine-tune the model with Instruction-following format with input, output pairs for **supervised fine-tuning** tasks. The basic structure includes:

Expand Down Expand Up @@ -67,7 +58,7 @@ This data teaches the model to adapt its behavior to suit your **specific domain
```
**Samples**: https://github.com/fpt-corp/ai-studio-samples/tree/main/sample-datasets/alpaca

## **2. ShareGPT**
### **2.2 ShareGPT**

**a. Trainer = SFT**

Expand Down Expand Up @@ -126,7 +117,7 @@ Each data sample consists of a `conversations` array, where each turn in the cha

**Samples:** https://github.com/fpt-corp/ai-studio-samples/tree/main/sample-datasets/sharegpt

## **4. ShareGPT_Image**
### **2.3 ShareGPT_Image**

**ShareGPT_Image** is an extension of the ShareGPT multi-turn chat format, designed specifically for multi-modal training — that is, training models that handle both text and images in conversations.

Expand Down Expand Up @@ -216,31 +207,30 @@ The structure includes:

**Samples**: [https://github.com/fpt-corp/ai-studio-samples/tree/main/sample-datasets/sharegpt-image](https://github.com/fpt-corp/ai-studio-samples/tree/main/sample-datasets/sharegpt-image)

## 2. Training and Validation Data
## 3. Training and Validation Data

- **Training data (required):** the main dataset used for model training.
- **Validation data (recommended):** used to evaluate model quality during training.

### 2.1. Data Split Rules
### 3.1. Data Split Rules
- **Train/Validation split:** 80% / 20%.
- **Small dataset (<2,000 samples):** you may use the entire dataset for training, but quality will be harder to verify.
- **Large dataset (>10,000 samples):** always prepare a separate validation set.

---


## 3. Dataset Validation Rules
## 4. Dataset Validation Rules

In addition to checking file format (CSV, JSON, JSONL, Parquet, ZIP), the system should also validate the dataset content before fine-tuning.


### 3.1. Basic Validation (Format-level)
### 4.1. Basic Validation (Format-level)
- File size must not exceed **100MB**.
- Must be in a supported format.
- Training set must contain at least **1,000 samples**.


### 3.2. Content Validation (Content-level)
### 4.2. Content Validation (Content-level)
**Structure:**
- Each record must contain **2 fields: `prompt` and `completion`**.
- Records with missing or empty fields are not accepted.
Expand All @@ -262,14 +252,12 @@ In addition to checking file format (CSV, JSON, JSONL, Parquet, ZIP), the system
- Prompts and completions should be well-distributed, not biased toward a single type of question.


### 3.3. Advanced Validation (Recommended)
### 4.3. Advanced Validation (Recommended)
- Check for dataset bias (no harmful or sensitive content).
- Ensure **UTF-8 encoding** to avoid parsing errors.

---


## 4. Common Issues & How to Avoid Them
## 5. Common Issues & How to Avoid Them
- **Context drift:** Always keep prompt formatting consistent.
- **Overly long or redundant completions:** Keep only the necessary output.
- **Noisy data (typos, duplicates):** Clean before uploading.
Expand Down