Examples & Tutorials

Guides to train with Jobs

Guides for using popular libraries with Jobs:

• Training with TRL on Jobs - Run SFT, GRPO, DPO and more using TRL and TRL Jobs
• Fine-tune with Unsloth on Jobs - ~2x faster training and ~60% less VRAM using Unsloth
• Transformers example scripts - UV-compatible training scripts for text classification, summarization, image classification, NER, speech recognition, and more — run directly on Jobs:

hf jobs uv run --flavor a10g-small --secrets HF_TOKEN \
  https://raw.githubusercontent.com/huggingface/transformers/main/examples/pytorch/image-classification/run_image_classification.py \
  --model_name_or_path google/vit-base-patch16-224-in21k \
  --dataset_name ethz/food101 \
  --output_dir vit-food101 \
  --push_to_hub

Process data at scale

DataTrove provides an experimental JobsPipelineExecutor for distributing data processing pipelines across a pool of Jobs. It supports concurrency limits, multi-stage dependencies, retries, and resumable runs — re-running a pipeline skips tasks that already completed and only runs the remaining ones.

See the ready-to-run examples for:

• Filtering a Hub dataset
• Tokenizing and merging a Hub dataset
• Multi-stage MinHash deduplication

UV Scripts

The uv-scripts organization maintains a collection of self-contained uv scripts that run on Jobs with a single command. Scripts cover OCR, batch inference, text classification, object detection, dataset statistics, embedding visualization, and more.

Unsloth also provides ready-to-run training scripts for fine-tuning LLMs and VLMs on Jobs.

Coding Agent Skills

Coding agents like Claude Code, Codex and Cursor can submit and monitor Jobs for you. Install the hf CLI skill, generated from your installed CLI so it stays current:

hf skills add

See Hugging Face CLI for AI agents for setup per agent, and Agent Skills for training and other workflow skills.

Sandboxes

The expose ports feature of Jobs makes them a great fit for building sandboxes, i.e. temporary self-contained environments used by agents and LLM applications.

Community Tutorials and Projects

• Train on massive datasets without downloading - Stream datasets directly on Jobs with Unsloth, no local storage needed
• Fine-tune a vision-language model with TRL - Fine-tune Qwen2.5-VL for art history tasks using TRL and Jobs
• FreeFlow - Open-source annotation platform with built-in Jobs integration for training YOLOv11 object detection models

Have a tutorial or project using Jobs? Open a PR to add it here.