Skip to content

HuggingFace Provider

The HuggingFace provider is the default, standard provider for ModelForge with maximum compatibility.

Overview

The HuggingFace provider uses the standard Transformers library to load and fine-tune models from the HuggingFace Hub.

Features

✅ Maximum compatibility - Works with all HuggingFace models
✅ All platforms - Windows, Linux
✅ Well-tested - Battle-tested implementation
✅ Extensive documentation - Full HuggingFace ecosystem support
✅ All tasks - Text generation, summarization, question answering

Installation

The HuggingFace provider is included by default with ModelForge:

pip install modelforge-finetuning

Usage

Basic Configuration

{
  "provider": "huggingface",
  "model_name": "meta-llama/Llama-3.2-3B",
  "task": "text-generation",
  "strategy": "sft"
}

Via UI

The HuggingFace provider is selected by default in the UI.

Via API

curl -X POST http://localhost:8000/api/start_training \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "huggingface",
    "model_name": "meta-llama/Llama-3.2-3B",
    "task": "text-generation"
  }'

Supported Models

All models compatible with HuggingFace Transformers:

  • Llama (all versions)
  • Mistral (all versions)
  • Qwen (all versions)
  • Gemma (all versions)
  • Phi (all versions)
  • BART
  • T5
  • BERT-based models
  • And thousands more!

Comparison with Unsloth

Feature HuggingFace Unsloth
Training Speed 1x (baseline) 2x faster
Memory Usage Baseline -20%
Platform Support All Linux/WSL/Docker only
Model Support All Limited to specific architectures
Stability Very Stable Stable
Maturity Mature Newer

When to Use HuggingFace

Use HuggingFace provider when:

✅ Running on native Windows
✅ Using models not supported by Unsloth (BART, T5, etc.)
✅ Maximum compatibility is important
✅ You need the most stable option
✅ First time using ModelForge

Configuration Tips

Optimal Settings

{
  "provider": "huggingface",
  "use_4bit": true,
  "bf16": true,
  "gradient_checkpointing": true,
  "per_device_train_batch_size": 4,
  "gradient_accumulation_steps": 4
}

For Low VRAM

{
  "provider": "huggingface",
  "strategy": "qlora",
  "use_4bit": true,
  "per_device_train_batch_size": 1,
  "gradient_accumulation_steps": 16,
  "gradient_checkpointing": true
}

Next Steps

HuggingFace Provider: Reliable and compatible! 🤗