text large_stringlengths 113 3.16k |
|---|
<s>[INST] What is Unsloth and what can it do? [/INST] Unsloth lets you run and train AI models on your own local hardware via an open-source UI. Our docs will guide you through running & training your own LLM locally. It supports inference and training for 500+ models across text, audio, embedding, vision, and more. Ke... |
<s>[INST] How do I install Unsloth on Windows? [/INST] To install Unsloth on Windows, open **Windows PowerShell** and run:
```
irm https://unsloth.ai/install.ps1 | iex
``` </s> |
<s>[INST] What are the benefits of fine-tuning an LLM? [/INST] **Fine-tuning** customizes an LLM’s behavior, enhances domain knowledge, and optimizes performance for specific tasks. Benefits include:
- **Update Knowledge**: Introduce new domain-specific information.
- **Customize Behavior**: Adjust tone, personality, ... |
<s>[INST] How does reinforcement learning (RL) work in Unsloth? [/INST] **Reinforcement Learning (RL)** in Unsloth involves an "agent" (the model) learning to make decisions by interacting with an environment and receiving feedback via **rewards** or **penalties**. Key components:
- **Action**: What the model generate... |
<s>[INST] Which models does Unsloth support for local training and inference? [/INST] Unsloth supports **500+ models**, including:
- **Text models**: GLM-5.2, DeepSeek-V4, Qwen3.6, Gemma 4 (with ~2x speedup), Llama 4, Mistral, Phi-4, and more.
- **Specialized models**: Vision, TTS (text-to-speech), embedding, and RL.
... |
<s>[INST] How do I install Unsloth on MacOS/Linux/WSL? [/INST] Run the following command in your terminal:
```
curl -fsSL https://unsloth.ai/install.sh | sh
```
For **Docker**, use the official image: [`unsloth/unsloth`](https://hub.docker.com/r/unsloth/unsloth) (supports Windows, WSL, and Linux; MacOS coming soon). <... |
<s>[INST] What are the key differences between fine-tuning and RAG? [/INST] **Fine-tuning** customizes a pre-trained model (e.g., Llama-3.1-8B) to perform specific tasks by training it on domain-specific data. It can replicate **all of RAG’s capabilities** (e.g., domain knowledge, task optimization) but is **not limite... |
<s>[INST] How can I use Unsloth’s API for inference? [/INST] Unsloth provides an **API endpoint** for running inference. You can:
1. Use **Unsloth Studio**’s built-in API tools (see [Unsloth API docs](/docs/basics/api)).
2. Integrate with external applications via HTTP requests.
3. Leverage **auto-healing tool calling... |
<s>[INST] What are the system requirements for Unsloth? [/INST] Unsloth supports **MacOS, Linux, Windows, NVIDIA, Intel, and CPU setups**. For detailed requirements (e.g., GPU VRAM, dependencies), see the [Unsloth Requirements guide](/docs/get-started/fine-tuning-for-beginners/unsloth-requirements). </s> |
<s>[INST] How does Unsloth improve model accuracy compared to other tools? [/INST] Unsloth **directly collaborates** with teams behind models like **gpt-oss, Qwen3, Llama 4, Mistral, Gemma, and Phi-4**, fixing **critical bugs** that improve accuracy. For example:
- **Qwen3**: Fixed dynamic GGUF bugs for 128K context s... |
<s>[INST] Can Unsloth handle multi-GPU training? What’s the status? [/INST] Unsloth **supports multi-GPU training**, but a **better version is coming soon**. For current implementations, see the [Multi-GPU Training Guide](/docs/basics/multi-gpu-training-with-unsloth). </s> |
<s>[INST] What is Unsloth Studio, and how is it different from the CLI? [/INST] **Unsloth Studio** is a **new open, no-code UI** for training and running LLMs, offering:
- **Visual workflows** for dataset creation (PDF/CSV/DOCX → training-ready data).
- **Live observability**: Monitor training (loss, GPU usage, custom... |
<s>[INST] I'm completely new to fine-tuning LLMs. What are the first steps I should take? [/INST] If you're a beginner, start by checking out the **Fine-tuning Guide** for a step-by-step walkthrough:
🧬 [Fine-tuning Guide](/docs/get-started/fine-tuning-llms-guide)
This guide covers the core basics of training. You ca... |
<s>[INST] How do I decide between an Instruct or Base Model for fine-tuning? [/INST] To choose between an Instruct or Base Model, refer to the **What Model Should I Use?** guide:
❓ [What Model Should I Use?](/docs/get-started/fine-tuning-llms-guide/what-model-should-i-use)
This guide also covers:
- How big your datas... |
<s>[INST] I want to fine-tune DeepSeek. What are the recommended settings? [/INST] For guidance on running and fine-tuning **DeepSeek**, check the **Complete LLM Directory**:
🚀 [Complete LLM Directory](/docs/models/tutorials)
This includes:
- Step-by-step instructions for DeepSeek.
- Model-specific settings (e.g., f... |
<s>[INST] How do I prepare and structure my dataset for fine-tuning? [/INST] To structure and prepare your dataset, follow the **Datasets Guide**:
📈 [Datasets Guide](/docs/get-started/fine-tuning-llms-guide/datasets-guide)
This covers:
- Dataset formatting requirements.
- Data collection best practices.
For hardwar... |
<s>[INST] What VRAM do I need to run Unsloth, and will it work on my GPU? [/INST] Check **Unsloth Requirements** for GPU compatibility and VRAM needs:
🛠️ [Unsloth Requirements](/docs/get-started/fine-tuning-for-beginners/unsloth-requirements)
This guide answers:
- Minimum VRAM requirements.
- Supported GPU types (e.... |
<s>[INST] How do I save my fine-tuned model locally and deploy it via Ollama or vLLM? [/INST] For saving and deploying your model:
🖥️ [Inference & Deployment](/docs/basics/inference-and-deployment)
This includes:
- Steps to save your model locally.
- Instructions for running via **Ollama** or **vLLM**. </s> |
<s>[INST] What happens if I change hyperparameters during fine-tuning? How should I adjust them? [/INST] To understand hyperparameter effects and recommended adjustments, see the **Hyperparameters Guide**:
🧠 [Hyperparameters Guide](/docs/get-started/fine-tuning-llms-guide/lora-hyperparameters-guide)
This explains:
-... |
<s>[INST] Where can I find the complete documentation for Unsloth models? [/INST] For the complete documentation index, see [llms.txt](https://unsloth.ai/docs/llms.txt). This page is also available as [Markdown](https://unsloth.ai/docs/get-started/unsloth-model-catalog.md). </s> |
<s>[INST] What are the different model formats available for Unsloth LLMs? [/INST] Unsloth LLMs directory supports the following formats:
• GGUF + 4-bit
• Instruct 16-bit
• Base 4 & 16-bit
• FP8 </s> |
<s>[INST] Which tools can I use to run GGUF models? [/INST] **GGUFs** let you run models in tools like [**Unsloth Studio**](/docs/new/studio)✨, Ollama, and llama.cpp. </s> |
<s>[INST] What is the recommended Qwen3.6 model variant and where can I find its GGUF files? [/INST] The recommended **Qwen3.6** model variant is **27B** with GGUF files available at:
[link](https://huggingface.co/unsloth/Qwen3.6-27B-GGUF) • [MTP](https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF)
For the **35B-A3B*... |
<s>[INST] How can I access the Qwen3.5 model in different sizes and where are the GGUF files hosted? [/INST] The **Qwen3.5** model is available in multiple sizes with GGUF files hosted at:
• 35B-A3B: [link](https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF)
• 27B: [link](https://huggingface.co/unsloth/Qwen3.5-27B-GGU... |
<s>[INST] Where can I find the Hugging Face links for Mistral 3 models in different sizes? [/INST] The **Mistral 3** models are available in the following sizes with Hugging Face links:
• 3B:
- Instruct: [link](https://huggingface.co/unsloth/Ministral-3-3B-Instruct-2512-GGUF)
- Reasoning: [link](https://huggingface... |
<s>[INST] What are the available Qwen3-VL model variants and where can I find their GGUF files? [/INST] The **Qwen3-VL** model is available in the following variants with GGUF files hosted at:
• 2B-Instruct:
- [link](https://huggingface.co/unsloth/Qwen3-VL-2B-Instruct-GGUF)
- 4-bit: [link](https://huggingface.co/un... |
<s>[INST] What are the system requirements for using Unsloth Studio on Windows? [/INST] Unsloth Studio works directly on Windows without WSL. To train models, your system must meet these requirements:
- Windows 10 or Windows 11 (64-bit)
- NVIDIA GPU with drivers installed
- **App Installer** (includes `winget`): [here... |
<s>[INST] Can I use Unsloth Studio on my Mac for training models? What are the requirements? [/INST] Unsloth Studio works on Mac devices for **Chat** for GGUF models and **Data Recipes** (Export coming very soon). **MLX training is coming soon!**
Requirements:
- macOS 12 Monterey or newer (Intel or Apple Silicon)
- In... |
<s>[INST] What are the GPU VRAM requirements for fine-tuning a 7B parameter model using QLoRA vs. LoRA? [/INST] For a **7B parameter model**, the VRAM requirements are:
- **QLoRA (4-bit)**: 5 GB
- **LoRA (16-bit)**: 19 GB
*Note: These are the absolute minimum requirements. Sometimes more VRAM may be needed depending ... |
<s>[INST] How do I set up Unsloth Core on Linux? What are the key requirements? [/INST] Unsloth Core works on **Linux** with these key requirements:
- **Operating System**: Linux (Ubuntu 20.04+ recommended)
- **GPU Support**: NVIDIA GPUs (2018+), including Blackwell RTX 50 and DGX Spark
- **CUDA Capability**: Minimum ... |
<s>[INST] What are the minimum VRAM requirements for fine-tuning a 405B parameter model? [/INST] For a **405B parameter model**, the VRAM requirements are:
- **QLoRA (4-bit)**: 237 GB
- **LoRA (16-bit)**: 950 GB
*Note: These are the absolute minimum requirements. Adjust batch size to 1, 2, or 3 to reduce VRAM usage.*... |
<s>[INST] Can I use Unsloth Studio on a CPU-only system? What features are supported? [/INST] Unsloth Studio supports **CPU-only devices** for:
- **Chat** for GGUF models
- **Data Recipes** (Export feature coming soon)
*Note: Training is not supported on CPU-only systems. Ensure your system meets the same requirement... |
<s>[INST] What are the differences between Unsloth Studio and Unsloth Core in terms of GPU support? [/INST] Unsloth Studio currently supports **NVIDIA GPUs** for training, with **AMD, MLX, and Intel support coming soon**. For **AMD and Intel GPU training**, you must use the **original Unsloth Core**, which supports:
-... |
<s>[INST] How do I install the required dependencies for Unsloth Core on Linux? [/INST] To install dependencies for **Unsloth Core** on Linux, ensure you have:
- **Git**: Usually preinstalled or install via `sudo apt install git`
- **CMake**: Preinstalled or install via `sudo apt install cmake`
- **C++ compiler**: Ins... |
<s>[INST] What is fine-tuning in the context of LLMs, and what are its main benefits over using a base model? [/INST] Fine-tuning an LLM customizes its behavior, deepens its domain expertise, and optimizes its performance for specific tasks. By refining a pre-trained model (e.g., *Llama-3.1-8B*) with specialized data, ... |
<s>[INST] Can fine-tuning add new knowledge to a model, or does it only adjust existing knowledge? [/INST] **Yes, fine-tuning can absolutely add new knowledge to a model.** A common misconception suggests it only adjusts existing knowledge, but if your fine-tuning dataset contains new domain-specific information, the m... |
<s>[INST] How does fine-tuning compare to Retrieval-Augmented Generation (RAG) in terms of capabilities? [/INST] Fine-tuning can do most things RAG can, but not the other way around. Fine-tuning embeds external knowledge directly into the model, allowing it to handle niche queries, summarize documents, and maintain con... |
<s>[INST] Is fine-tuning expensive, and what are some cost-effective alternatives? [/INST] **Not at all!** While full fine-tuning or pretraining can be costly, these are unnecessary for most use cases. Instead, techniques like **LoRA (Low-Rank Adaptation)** or **QLoRA (Quantized LoRA)** can be used for minimal cost. Wi... |
<s>[INST] Why should I combine RAG and fine-tuning instead of choosing just one approach? [/INST] Combining RAG and fine-tuning leverages the strengths of both approaches for the best results:
- **Task-Specific Expertise**: Fine-tuning excels at specialized tasks or formats, making the model an expert in a specific ar... |
<s>[INST] What are some real-world applications of fine-tuning LLMs? [/INST] Fine-tuning can be applied across various domains, including:
- **Sentiment Analysis for Finance**: Train an LLM to determine if a news headline impacts a company positively or negatively, tailoring its understanding to financial context.
- ... |
<s>[INST] Where can I find Unsloth's Colab notebooks for training models like Gemma 4 or Qwen3.5? [/INST] You can find Unsloth's Colab notebooks for training models like **Gemma 4** (E4B Vision, E2B Text, E2B Audio, 31B on Kaggle) and **Qwen3.5** (0.8B, 2B, 4B Vision) directly from their links:
- **Gemma 4**:
- [E4B... |
<s>[INST] How do I access the Unsloth Studio Colab notebook for training models under 22B parameters? [/INST] You can access the **Unsloth Studio** Colab notebook for training models under 22B parameters by clicking the following link:
[Unsloth Studi... |
<s>[INST] Which notebooks are available for fine-tuning Qwen3 models, including multimodal capabilities? [/INST] The following notebooks are available for fine-tuning **Qwen3** models, including multimodal capabilities:
- **Qwen3 (14B) - Reasoning & Conversational**: [Notebook](https://colab.research.google.com/github... |
<s>[INST] Where can I find notebooks for reinforcement learning (GRPO) tasks like Sudoku? [/INST] You can find notebooks for **GRPO (Reasoning RL)** tasks, including Sudoku, for the following models:
- **Gemma 4 E2B - Reinforcement Learning (Sudoku Game)**: [Notebook](https://colab.research.google.com/github/unslothai... |
<s>[INST] What are the available notebooks for training and inference with gpt-oss (20B)? [/INST] The following notebooks are available for **gpt-oss (20B)**:
- **Fine-tuning**: [Notebook](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/gpt-oss-(20B)-Fine-tuning.ipynb)
- **Inference**: [Noteb... |
<s>[INST] How can I access notebooks for training models like Gemma 3, including multimodal and function-specific variants? [/INST] You can access notebooks for training **Gemma 3** models, including multimodal and function-specific variants, via the following links:
- **Gemma 3 (4B)**: [Notebook](https://colab.resear... |
<s>[INST] Where can I find the Unsloth model catalog with GGUF, 4-bit, and 16-bit models? [/INST] Unsloth LLMs directory for all our [Dynamic arrow-up-right](https://docs.unsloth.ai/basics/unsloth-dynamic-2.0-ggufs) GGUF, 4-bit, 16-bit models on Hugging Face. The catalog includes:
• GGUF + 4-bit
• Instruct 16-bit
• Bas... |
<s>[INST] What tools can I use to run GGUF models? [/INST] GGUFs let you run models in tools like [**Unsloth Studio**](/docs/new/studio) ✨, Ollama, and llama.cpp. </s> |
<s>[INST] How do I access the Qwen3.6-27B model in GGUF format? [/INST] You can access the Qwen3.6-27B model in GGUF format via this Hugging Face link: [link arrow-up-right](https://huggingface.co/unsloth/Qwen3.6-27B-GGUF). </s> |
<s>[INST] Where can I find the 4-bit Instruct version of the Gemma 4-31B model? [/INST] The 4-bit Instruct version of the Gemma 4-31B model is available at this Hugging Face link: [link arrow-up-right](https://huggingface.co/unsloth/gemma-4-31B-it-unsloth-bnb-4bit). </s> |
<s>[INST] What are the available variants of the Gemma 4 model in GGUF format? [/INST] The available variants of the Gemma 4 model in GGUF format are:
• 26B-A4B: [link arrow-up-right](https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF)
• 31B: [link arrow-up-right](https://huggingface.co/unsloth/gemma-4-31B-it-GGUF)... |
<s>[INST] Can you provide the Hugging Face link for the Kimi K2.6 model in GGUF format? [/INST] The Hugging Face link for the Kimi K2.6 model in GGUF format is: [link arrow-up-right](https://huggingface.co/unsloth/Kimi-K2.6-GGUF). </s> |
<s>[INST] What are the two ways to use Unsloth and how do they differ? [/INST] Unsloth can be used in two ways: through **Unsloth Studio**, the web UI, or through **Unsloth Core**, the original code-based version. Unsloth Studio is a graphical interface accessible via a web browser, while Unsloth Core is the command-li... |
<s>[INST] How do I install Unsloth on MacOS or Linux? [/INST] To install Unsloth on MacOS or Linux, run the following command:
```
curl -fsSL https://unsloth.ai/install.sh | sh
``` </s> |
<s>[INST] What is the command to install Unsloth on Windows? [/INST] To install Unsloth on Windows using PowerShell, run:
```
irm https://unsloth.ai/install.ps1 | iex
``` </s> |
<s>[INST] How do I launch Unsloth Studio after installation? [/INST] To launch Unsloth Studio, use the following command:
```
unsloth studio -H 0.0.0.0 -p 8888
``` </s> |
<s>[INST] Where can I find detailed installation instructions for specific platforms like AMD, Intel, or Conda? [/INST] Detailed installation instructions for specific platforms are available at the following links:
- [AMD](/docs/get-started/install/amd)
- [Intel](/docs/get-started/install/intel)
- [Conda](/docs/get-st... |
<s>[INST] How do I update Unsloth after installation? [/INST] To update Unsloth, you can use the same installation command as before or run:
```
unsloth studio update
``` </s> |
<s>[INST] How do I install Unsloth Studio on macOS using the command line? [/INST] To install Unsloth Studio on macOS, use the following command:
```
curl -fsSL https://unsloth.ai/install.sh | sh
```
To update or launch Unsloth Studio, use:
```
unsloth studio update
unsloth studio -H 0.0.0.0 -p 8888
``` </s> |
<s>[INST] What is the command to install Unsloth Studio on Windows using PowerShell? [/INST] To install Unsloth Studio on Windows using PowerShell, run:
```
irm https://unsloth.ai/install.ps1 | iex
```
To update or launch Unsloth Studio, use:
```
unsloth studio update
unsloth studio -H 0.0.0.0 -p 8888
``` </s> |
<s>[INST] How do I install Unsloth from the main repository on Linux? [/INST] To install Unsloth from the main repository on Linux, follow these steps:
```
git clone https://github.com/unslothai/unsloth
cd unsloth
./install.sh --local
unsloth studio -H 0.0.0.0 -p 8888
``` </s> |
<s>[INST] What are the steps to uninstall Unsloth Studio completely on Windows? [/INST] To uninstall Unsloth Studio completely on Windows, follow these 4 steps:
1. Remove the application:
```
Remove-Item -Recurse -Force "$HOME\.unsloth\studio\unsloth_studio"
```
2. Remove shortcuts and symlinks:
```
Remove-Item -Forc... |
<s>[INST] How can I install Unsloth Core using uv pip for the latest release? [/INST] To install Unsloth Core using uv pip for the latest release, follow these steps:
```
curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv unsloth_env --python 3.13
source unsloth_env/bin/activate
uv pip install unsloth --torch-bac... |
<s>[INST] What is the correct command to install Unsloth with vLLM on a system with CUDA 12.1 and torch 2.4? [/INST] To install Unsloth with vLLM on a system with CUDA 12.1 and torch 2.4, use:
```
uv pip install unsloth vllm --torch-backend=auto
```
If you need to install Unsloth with specific CUDA and torch versions... |
<s>[INST] How do I delete cached Hugging Face model files on Windows? [/INST] To delete cached Hugging Face model files on Windows, navigate to the default cache directory:
```
%USERPROFILE%\.cache\huggingface\hub\
```
If `HF_HUB_CACHE` or `HF_HOME` is set, use that location instead. You can manually delete the relev... |
<s>[INST] What are the steps to install Unsloth Nightly on Linux? [/INST] To install Unsloth Nightly on Linux, follow these steps:
```
git clone https://github.com/unslothai/unsloth
cd unsloth
git checkout nightly
./install.sh --local
```
Then to launch every time:
```
unsloth studio -H 0.0.0.0 -p 8888
``` </s> |
<s>[INST] How do I force reinstall Unsloth Core if I'm experiencing dependency issues? [/INST] To force reinstall Unsloth Core and resolve dependency issues, use:
```
pip install --upgrade --force-reinstall --no-cache-dir --no-deps unsloth
pip install --upgrade --force-reinstall --no-cache-dir --no-deps unsloth_zoo
``... |
<s>[INST] What is the command to install Unsloth on macOS? [/INST] To install Unsloth locally on macOS, use the following command:
```
curl -fsSL https://unsloth.ai/install.sh | sh
``` </s> |
<s>[INST] How do I update Unsloth to the latest version? [/INST] You can update Unsloth using either of these commands:
1. The same installation command:
```
curl -fsSL https://unsloth.ai/install.sh | sh
```
2. Using the Unsloth Studio update command:
```
unsloth studio update
``` </s> |
<s>[INST] What command should I use to launch Unsloth Studio on macOS? [/INST] To launch Unsloth Studio, use the following command:
```
unsloth studio -H 0.0.0.0 -p 8888
``` </s> |
<s>[INST] How do I completely uninstall Unsloth Studio from macOS, including all data? [/INST] To uninstall Unsloth Studio and **all** associated data (history, cache, chats, model checkpoints, and exports), follow these steps:
1. Remove the application:
```
rm -rf ~/.unsloth
```
2. Remove shortcuts and symlinks:
```... |
<s>[INST] Where are Hugging Face model files stored on macOS, and how can I delete them? [/INST] Hugging Face model files are stored in the default cache location:
```
~/.cache/huggingface/hub/
```
To check the exact cache location (in case of custom environment variables), run:
```
echo ${HF_HUB_CACHE:-${HF_HOME:-$... |
<s>[INST] What is the difference between uninstalling Unsloth Studio and removing everything? [/INST] When you uninstall Unsloth Studio using:
```
rm -rf ~/.unsloth/studio/unsloth_studio
```
It **only removes the application** while keeping:
- Model checkpoints
- Exports
- History
- Cache
- Chats
However, if you run:... |
<s>[INST] How can I fine-tune a large language model like Llama 70B using multiple GPUs with Unsloth? [/INST] Unsloth currently supports multi-GPU setups through libraries like Accelerate and DeepSpeed, allowing you to use parallelism methods such as **FSDP** and **DDP**. For now, you can use our [Magistral-2509 Kaggle... |
<s>[INST] What command should I use to launch multi-GPU training with Unsloth? [/INST] To enable multi-GPU training for DDP, run either:
`accelerate launch train.py`
or
`torchrun --nproc_per_node N_GPUS train.py`
where `N_GPUS` is the number of GPUs you have. </s> |
<s>[INST] How can I load a 70B parameter model like Llama 3.3 on a single GPU with limited VRAM? [/INST] If you don't have enough VRAM for a single GPU to load a model like Llama 70B, Unsloth can split the model across GPUs using the `device_map = "balanced"` flag. Here's how:
```python
from unsloth import FastLanguag... |
<s>[INST] Where can I find an example of multi-GPU Unsloth training for a 24B parameter model? [/INST] You can refer to our [Magistral-2509 Kaggle notebook](https://github.com/unslothai/notebooks/blob/main/notebooks/Magistral-2509-Kaggle.ipynb), which demonstrates how to fine-tune a 24B parameter model using multi-GPU ... |
<s>[INST] What is the current status of Unsloth's official multi-GPU support? [/INST] Unsloth currently supports multi-GPU setups through libraries like Accelerate and DeepSpeed, but the process can be complex. We are actively working to simplify this and will announce official multi-GPU support soon. For now, you can ... |
<s>[INST] How can I get started with multi-GPU training using Unsloth? [/INST] To get started with multi-GPU training using Unsloth:
1. Create a training script (e.g., `train.py`) using one of our [training scripts](https://github.com/unslothai/notebooks/tree/main/python_scripts).
2. Run the script with either:
- `a... |
<s>[INST] How do I install Unsloth from source for multi-GPU training? [/INST] To install Unsloth from source, follow these steps:
1. Clone the Unsloth repository from GitHub:
```
git clone https://github.com/unslothai/unsloth.git
cd unsloth
```
2. Install it using pip:
```
pip install .
```
**Note:** Consider using... |
<s>[INST] What are the required and optional parameters for running Unsloth CLI for multi-GPU training? [/INST] For multi-GPU training using Unsloth CLI with Distributed Data Parallel (DDP), the **required** parameters are:
- `--model_name` (e.g., `Qwen/Qwen3-8B`)
- `--dataset` (e.g., `yahma/alpaca-cleaned`)
**Option... |
<s>[INST] How do I verify my GPUs are available before starting multi-GPU training with Unsloth? [/INST] To check your GPU status before training, use the `nvidia-smi` command-line tool. For example, if you have two H100 GPUs, the output should look like this:
```
nvidia-smi
```
Expected output (example):
```
+------... |
<s>[INST] What command should I use to start multi-GPU training with Unsloth on two H100 GPUs? [/INST] To start multi-GPU training with Unsloth on two H100 GPUs using Distributed Data Parallel (DDP), use the following command:
```
torchrun --nproc_per_node=2 unsloth-cli.py \
--model_name=Qwen/Qwen3-8B \
--dataset=yahm... |
<s>[INST] Can I use Unsloth DDP with my own training scripts, not just the CLI? [/INST] Yes, Unsloth DDP will work with **any** of your training scripts, not just the CLI. For example, you can use the `torchrun` launcher with scripts converted from Unsloth's free Colab notebooks (e.g., [scripts arrow-up-right](https://... |
<s>[INST] What is the purpose of Distributed Data Parallel (DDP) in Unsloth multi-GPU training? [/INST] Distributed Data Parallel (DDP) in Unsloth creates **one copy of the model on each GPU**, feeding each copy **distinct samples** from the dataset during training. This approach:
1. **Scales training throughput** by ... |
<s>[INST] What are the benefits of fine-tuning embedding models with Unsloth? [/INST] Fine-tuning embedding models with Unsloth improves retrieval and RAG performance on specific tasks by aligning the model's vectors with your domain and the kind of 'similarity' that matters for your use case. This reduces errors and i... |
<s>[INST] Which models does Unsloth support for fine-tuning, and what are their key use cases? [/INST] Unsloth supports fine-tuning for embedding, classifier, BERT, and reranker models. Key models and their use cases include:
- **EmbeddingGemma (300M)**: Lightweight embedding model.
- **Qwen3-Embedding (4B/0.6B)**: Hig... |
<s>[INST] How can I fine-tune an embedding model with Unsloth? Where can I find the notebooks? [/INST] Unsloth provides free fine-tuning notebooks for various models. You can access them directly via:
- [EmbeddingGemma (300M)](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/EmbeddingGemma_(300... |
<s>[INST] What are the key features of Unsloth for fine-tuning embedding models? [/INST] Unsloth offers the following key features:
- LoRA/QLoRA or full fine-tuning for embeddings without pipeline rewrites.
- Best support for encoder-only `SentenceTransformer` models (with `modules.json`).
- Cross-encoder models train ... |
<s>[INST] How do I load a fine-tuned Unsloth model for inference? What is the correct way to initialize it? [/INST] To load a fine-tuned Unsloth model for inference, use `FastSentenceTransformer.from_pretrained()` with the `for_inference=True` argument. Example:
```python
model = FastSentenceTransformer.from_pretraine... |
<s>[INST] How can I use my fine-tuned Unsloth model for semantic search or RAG? What tools are compatible? [/INST] Your fine-tuned Unsloth model can be deployed with major tools like:
- Transformers
- LangChain
- Weaviate
- SentenceTransformers
- Text Embeddings Inference (TEI)
- vLLM
- llama.cpp
- Custom embedding API... |
<s>[INST] What are the performance benchmarks for Unsloth compared to SentenceTransformers + Flash Attention 2? [/INST] Unsloth is consistently **1.8x to 3.3x faster** than SentenceTransformers + Flash Attention 2 (FA2) for 4bit QLoRA across various embedding models and sequence lengths (128–2048+). For example, **Embe... |
<s>[INST] What are the key benefits of fine-tuning TTS models instead of using zero-shot cloning? [/INST] Fine-tuning TTS models delivers far more accurate and realistic voice replication compared to zero-shot cloning. Zero-shot cloning captures tone but often misses pacing and expression, resulting in robotic and unna... |
<s>[INST] Which TTS models are supported by Unsloth for fine-tuning, and where can I find the corresponding notebooks? [/INST] Unsloth supports **any** `transformers`-compatible TTS model, including:
- [Sesame-CSM (1B)](https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Sesame_CSM_(1B)-TTS.ipynb)... |
<s>[INST] How do I load the Orpheus-TTS (3B) model using Unsloth with LoRA 16-bit? [/INST] To load the Orpheus-TTS (3B) model in LoRA 16-bit, use the following code:
```python
from unsloth import FastModel
model_name = "unsloth/orpheus-3b-0.1-pretrained"
model, tokenizer = FastModel.from_pretrained(
model_name,
... |
<s>[INST] What is the difference between Sesame-CSM (1B) and Orpheus-TTS (3B), and which one should I choose for fine-tuning? [/INST] **Sesame-CSM (1B)** is a base model requiring audio context for each speaker to perform well, while **Orpheus-TTS (3B)** is pre-trained on 8 professional voice actors and has built-in vo... |
<s>[INST] How do I prepare the Elise dataset for fine-tuning a TTS model with Unsloth? [/INST] The Elise dataset requires audio clips and corresponding transcripts. You can load it using Hugging Face's `datasets` library:
```python
from datasets import load_dataset, Audio
# Load the Elise dataset (e.g., with emotion ... |
<s>[INST] What should I do if my TTS output is limited to 10 seconds? [/INST] If the output duration is capped at 10 seconds, increase `max_new_tokens` from its default value of 125 (which corresponds to 10 seconds). For longer outputs, set a higher value, e.g., `max_new_tokens = 250` for 20 seconds.
**Example:**
```p... |
<s>[INST] What are the key improvements in Unsloth Dynamic 2.0 quantization compared to the previous version? [/INST] The key improvements in Unsloth Dynamic 2.0 quantization include:
- **Revamped Layer Selection for GGUFs + safetensors:** More intelligent and extensive layer quantization, adjusting quantization type ... |
<s>[INST] How does Unsloth Dynamic 2.0 perform in benchmarks compared to other quantization methods? [/INST] Unsloth Dynamic 2.0 outperforms leading quantization methods in multiple benchmarks:
- **Aider Polyglot:** Dynamic 3-bit DeepSeek V3.1 GGUF scores **75.6%**, surpassing many full-precision SOTA LLMs.
- **Qwen3.... |
End of preview. Expand in Data Studio
No dataset card yet
- Downloads last month
- 14