Spaces:
Sleeping
title: SLM Instruction Tuning using Unsloth
emoji: 🐨
colorFrom: yellow
colorTo: indigo
sdk: gradio
sdk_version: 4.38.1
app_file: app.py
pinned: false
license: apache-2.0
LoRA Instruction Tuning using Unsloth
What?
This Gradio app is a simple interface to access unsloth AI's fine-tuning methods but leveraging the A100 GPUs provided by Huggingface Spaces. This outputs of this fine-tuning will be instruction tuned LoRA weights that will be uploaded into your personal huggingface models repository.
Why?
The goal of this demo is to show how you can tune your own language models leveraging industry standard compute and fine tuning methods using a simple point-and-click UI.
In addition, compute, even on Google Colab's free tier is tight even with a T4 and rate limits are uncertain. This makes the use of the A100s on this demo useful for a small added boost to compute performance. For those looking to reduce the costs associated with training datasets can pull down the spaces repository to train their models at speed for $9 on The Huggingface Pro License.
This is a demo and not a production application. This application is subject a demand queue.
How?
Just start by following the guide below:
- Flip to the Train Model tab at the top.
- Populate your username, repository, token, and model details
- Upload data from transformers or your local jsonl file. Please view this guide for best practices.
- Eat a snack and wait as you train the model.
Coming soon!
- More models and added flexibility with guardrails on hyperparameter tuning.
- Downloads for a WandB logger for training monitoring.
Other resources.
- Unsloth's notebooks to look at what is going on under the hood.
