huggingface

name	huggingface
description	Onur Solmaz's personal Hugging Face tips and tricks. Use for any Hugging Face task alongside the relevant official Hugging Face marketplace skills, especially when personal operating conventions, deployment lessons, or recovery notes may apply. This skill supplements rather than replaces Hugging Face's published skills.

Hugging Face

Treat this as Onur's personal overlay for Hugging Face work. Apply it together with the relevant Hugging Face-published skills. Keep official product workflows in the official skills and keep only Onur's additional preferences and learned operational remedies here.

Work alongside the published skills

Discover the current Hugging Face marketplace catalog instead of assuming a fixed list:

hf skills list --format json

Use every installed official skill that materially covers the task. Typical routing includes hf-cli for Hub operations and task-specific skills for Spaces, datasets, Gradio, papers, local models, training, evaluations, Trackio, ZeroGPU, tool building, and cloud deployment. The marketplace is the source of truth when names or coverage change.

For training plans, data splits, checkpoint reuse, refits, ablations, distillation, or model-selection decisions, also use ml-experiment-design. Official training skills explain how to run Jobs. They do not replace the compute and experiment-design review required before launching them.

Before launching, scaling, retrying, or automatically continuing any paid accelerator Job, use paid-compute-launch. A bounded experiment with a hard cumulative ceiling below $5 may proceed without asking. Count every related attempt and recovery Job toward the same ceiling. Use hf-job-control for every substantial Job as defined there. A plain hf jobs run launch is acceptable only when the work remains below every substantial-launch threshold and its results are cheap to reproduce.

Long batch inference and dataset generation must publish durable output chunks while they run. Hugging Face Job logs, Trackio metrics, and row counters are monitoring evidence, not recoverable output. Before a multi-Job launch, pass the real worker's pause-resume canary and present measured hardware cost. Obtain approval unless the complete bounded experiment stays below $5.

If a matching official skill is unavailable locally, install or update the published copy through hf skills; do not duplicate its contents here:

hf skills add SKILL_NAME
hf skills update SKILL_NAME

Choose project or global installation deliberately. Do not silently overwrite a locally modified skill with --force.

Personal tips and tricks

Use the official Git credential integration

For HTTPS Git operations against huggingface.co, prefer Hugging Face's supported credential workflow:

hf auth login --add-to-git-credential

When the intended token is already stored by the Hugging Face CLI, select it explicitly instead of logging in again:

hf auth switch --token-name TOKEN_NAME --add-to-git-credential

Before either command persists a token, identify the credential source and the configured Git credential helper that will receive it, then obtain the explicit approval required by the credential handling policy. If Git has no suitable credential helper, explain the supported choices rather than silently selecting one.

Do not create a repository-local inline credential helper, temporary askpass script, or authentication wrapper when the supported workflow is available, unless Onur explicitly requests that alternative.

Prefer greedy decoding for T5 generation

Use greedy decoding as the default for T5-family translation, distillation, and large batch generation. Do not select beam search merely because it is common in machine translation. Apply practical-significance to the comparison. Beam search needs paired evidence of a stable, practically meaningful quality gain that justifies its measured time and cost. Changing from greedy still requires Onur's explicit approval.

The Alman ByT5 Base pilot illustrates the threshold. Beam-4 improved student exact match by only 12 of 2,953 development cases, or 0.41 percentage points. Greedy had better chrF and generation was about 41% faster. That result does not justify beam-4 as the default.

Recover a Space deployment that is not picked up

Before changing state, inspect the Space and both log streams:

hf auth whoami
hf spaces info OWNER/SPACE --format agent
hf spaces logs OWNER/SPACE --build --tail 120 --format agent
hf spaces logs OWNER/SPACE --tail 200 --format agent
hf spaces wait OWNER/SPACE --timeout 10m --format agent

Use a restart as the recovery action when all of the following are true:

• the intended repository SHA is already present;
• the stage remains transitional, such as RUNNING_BUILDING or RUNNING_APP_STARTING;
• build logs remain at Build Queued, or startup logs remain empty, across repeated checks for several minutes; and
• no builder, image-pull, or application-startup error is active.

Restart the existing revision once:

hf spaces restart OWNER/SPACE --format agent
hf spaces wait OWNER/SPACE --timeout 10m --format agent

This can make Hugging Face pick up an already-queued build or startup. Do not push an empty commit, rewrite Space files, or repeatedly restart it. A restart briefly interrupts the Space, so perform it directly only when the user asked to deploy, operate, or recover that Space. For read-only diagnosis, report it as the recommended remedy.

Afterward, require RUNNING, the intended runtime SHA, and application-specific health. If the Space reaches BUILD_ERROR or RUNTIME_ERROR, stop restarting and diagnose the first concrete error in the corresponding log.

Distinguish Hub access from application failures

A private Space returns an unsigned 404 by design. Do not treat that alone as an application routing failure. Use an authenticated or signed probe, then separately verify the repository SHA, runtime SHA, stage, build logs, run logs, and application health.

Clean up Dev Mode

Do not enable Dev Mode merely to read public logs. It changes live runtime state, and SSH access also requires a registered Hugging Face SSH key. When an explicitly authorized diagnosis uses Dev Mode, disable it afterward and wait for normal mode to return to RUNNING:

hf spaces dev-mode OWNER/SPACE --stop --format agent
hf spaces wait OWNER/SPACE --timeout 10m --format agent

If the return to normal mode stalls at RUNNING_APP_STARTING with empty logs, apply the one-time restart remedy above.

Maintaining this overlay

Add a tip here only when Onur explicitly asks to remember it or a repeated workflow establishes it as a personal convention. Keep each tip concrete, actionable, and narrower than the official skill it supplements.