unsloth/jobs · Datasets

unsloth/jobs · Datasets at Hugging Face 🦥 Unsloth Training Scripts for HF Jobs UV scripts for fine-tuning LLMs and VLMs using Unsloth on HF Jobs(on-demand cloud GPUs). UV handles dependency installation automatically, so you can run these scripts directly without any local setup. These scripts can also be used or adapted by agents to train models for you. Prerequisites - A Hugging Face account - The HF CLI installed and authenticated (`hf auth login`) - A dataset on the Hub in the appropriate format (see format requirements below). A strong LLM agent can often convert your data into the right format if needed. Data Formats LLM Fine-tuning (SFT) Requires conversation data in ShareGPT or similar format: ``` { "messages": {"from": "human", "value": "What is the capital of France?"}, {"from": "gpt", "value": "The capital of France is Paris."} } ``` The script auto-converts common formats (ShareGPT, Alpaca, etc.) via`standardize_data_formats`. See mlabonne/FineTome-100k for a working dataset example. VLM Fine-tuning Requires`images` and`messages` columns: ``` { "images": <PIL.Image>, List of images "messages": { "role": "user", "content": {"type": "image"}, {"type": "text", "text": "What's in this image?"} }, { "role": "assistant", "content": {"type": "text", "text": "A golden retriever playing fetch in a park."} } } ``` See davanstrien/iconclass-vlm-sft for a working dataset example, and davanstrien/iconclass-vlm-qwen3-best for a model trained with these scripts. Continued Pretraining Any dataset with a text column: ``` {"text": "Your domain-specific text here..."} ``` Use`--text-column` if your column has a different name. Usage View available options for any script: ``` uv run https://huggingface.co/datasets/unsloth/