uv-script
ocr
OCR UV Scripts
Part of uv-scripts — self-contained UV scripts you run on Hugging Face Jobs in one command.
A model zoo of OCR scripts — one per model — that add a markdown column to an image dataset. Pick a model from the table below, point it at your dataset, and run it on a GPU with one command. A few recipes do structured extraction instead — image or text → JSON given a schema (see Structured extraction below). Two more companions sit alongside: pp-doclayout.py detects… See the full description on the dataset page: https://huggingface.co/datasets/uv-scripts/ocr.transcription
Transcription
Scripts for transcribing — and diarizing — audio files using HF Buckets and Jobs.
Quick Start
Scripts run directly from their Hub URL — no clone or local checkout needed:
# 1. Download audio from Internet Archive straight into a bucket
hf jobs uv run \
-v hf://buckets/user/audio-files:/output \
https://huggingface.co/datasets/uv-scripts/transcription/raw/main/download-ia.py \
SUSPENSE /output
# 2. Transcribe — audio bucket in, transcript… See the full description on the dataset page: https://huggingface.co/datasets/uv-scripts/transcription.ocr-demo
OCR demo: Food for Space Flight
Seven scanned pages from NASA's Food for Space Flight booklet, with headings,
columns, photographs, food lists and tables. This small dataset is an input for
trying OCR recipes and inspecting their results.
The images are PDF pages 3-9 (printed pages 2-8) of the original booklet.
The selection omits the reproduction disclaimer and dark cover. The complete
original PDF and a matching seven-page extract are in the
OCR demo Bucket.
Use as… See the full description on the dataset page: https://huggingface.co/datasets/uv-scripts/ocr-demo.object-detection
Object Detection Dataset Scripts
8 scripts to create, convert, review, validate, inspect, diff, and sample object detection datasets on the Hub. Supports 6 bbox formats — no setup required.
Start from nothing: falcon-perception.py generates a first-pass detection dataset for any class you can name, zero-shot, with no labelling and no training. The other six then convert, check, and measure it.
This repository is inspired by panlabel
Quick Start
Convert bounding… See the full description on the dataset page: https://huggingface.co/datasets/uv-scripts/object-detection.training
Streaming LLM Training with Unsloth
Train on massive datasets without downloading anything - data streams directly from the Hub.
🦥 Latin LLM Example
Teaches Qwen Latin using 1.47M texts from FineWeb-2, streamed directly from the Hub.
Blog post: Train on Massive Datasets Without Downloading
Quick Start
# Run on HF Jobs (recommended - 2x faster streaming)
hf jobs uv run latin-llm-streaming.py \
--flavor a100-large \
--timeout 2h \
--secrets HF_TOKEN \
--… See the full description on the dataset page: https://huggingface.co/datasets/uv-scripts/training.classification
Classification Scripts
Text classification on HF Jobs — both directions:
Script
What it does
train-classifier.py
Fine-tune an encoder into a classifier (default: LFM2.5-Encoder-350M) and push it to the Hub
train-setfit.py
Few-shot train a classifier from 8-64 labels per class with SetFit — runs on CPU or GPU
classify-dataset.py
Zero-shot classify a dataset with an instruction LLM (SmolLM3 + vLLM, structured outputs)
classify-dataset-sglang.py
Zero-shot variant… See the full description on the dataset page: https://huggingface.co/datasets/uv-scripts/classification.