Agent Skill : hugging-face-jobs

Run any workload on Hugging Face Jobs.

Use this skill when you want to run GPU/CPU workloads (batch inference, synthetic data generation, dataset stats, experiments) on Hugging Face Jobs, with correct token handling and result persistence back to the Hub.

Skill Name: hugging-face-jobs

Main Documentation: SKILL.md

Scripts Directory: scripts/

References Directory: references/

Overview

This skill focuses on running real workloads via Hugging Face Jobs. It includes reviewable UV candidate scripts and guides for authentication (HF tokens), secrets vs env vars, timeouts, hardware selection, and pushing results to the Hub.

Core Documentation

SKILL.md
SKILL.md

Complete skill documentation (how to submit jobs, tokens/secrets, timeouts, persistence, and how to use the bundled scripts)

References

token_usage.md
references/token_usage.md

Token best practices: secrets vs env, permissions, common errors (401/403), and secure patterns
hub_saving.md
references/hub_saving.md

How to persist results: push datasets/models/files to the Hub (ephemeral job filesystem)
hardware_guide.md
references/hardware_guide.md

Flavor selection guidance for CPU/GPU/TPU workloads
troubleshooting.md
references/troubleshooting.md

Common failure modes (timeouts, missing deps, OOM, auth) and fixes

Scripts

generate-responses.py
scripts/generate-responses.py

vLLM batch generation: load prompts/messages from a dataset, generate responses, push dataset + card to Hub
cot-self-instruct.py
scripts/cot-self-instruct.py

CoT Self-Instruct synthetic data generation (reasoning/instruction) + optional filtering, pushes dataset + card
finepdfs-stats.py
scripts/finepdfs-stats.py

Polars streaming stats over Hub parquet (finepdfs-edu); optional upload of computed stats to a dataset repo