ml-intern: Detailed Introduction

[ikaijua]

ml-intern: Detailed Introduction

A HuggingChat mode that researches ML tasks, writes scripts, and runs training, evaluation, and demos on Hugging Face.

• Audience: Hugging Face users who can evaluate ML results and want to delegate bounded experiments.
• Cost: Chat inference and Jobs/sandboxes are billed separately; each conversation starts with a $0 compute budget.

Table of Contents

Getting started

1. What it is and who it is for
2. Access and setup: your first session

Usage and configuration

3. A concrete task and how to check it
4. Costs: software, models, and execution
5. Configuration and extensions

Limits and references

6. Permissions, data, and limitations
7. Frequently asked questions
8. Official references

Use the contents to jump to each section. If an in-page link does not respond, open it in a new tab.

[ikaijua]

What it is and who it is for

In one sentence: A HuggingChat mode that researches ML tasks, writes scripts, and runs training, evaluation, and demos on Hugging Face.

Who it suits: Hugging Face users who can evaluate ML results and want to delegate bounded experiments.

What you can do:

• Reproduce a paper at small scale or finetune a model.
• Build datasets and retain results on the Hub.
• Run evaluations or deploy a Gradio Space.

Where to start: The old ml-intern repository is archived; development moved to Chat UI. Start in HuggingChat.

Cost: Chat inference and Jobs/sandboxes are billed separately; each conversation starts with a $0 compute budget.

[ikaijua]

Access and setup: your first session

The old ml-intern repository is archived; development moved to Chat UI. Start in HuggingChat.

Sign in and enable ML Intern before the first message of a new conversation. Check enabled Hub tools in MCP settings and available credits.

[ikaijua]

A concrete task and how to check it

Start with a small evaluation, specifying model, dataset, output location, and budget.

Example:

Evaluate a small subset and show the plan first. Report the split, evaluation script, failures, and result links. Stop before exceeding the budget.

Inspect script versions, Jobs, sandboxes, and artifacts in Resources. Follow Trackio metrics when present; use Resume after an interrupted turn.

[ikaijua]

Costs: software, models, and execution

The conversation compute budget covers its Jobs and sandboxes, not chat inference, other HF products, or resources launched by a Job itself.

Launches reserve worst-case hardware cost up to their timeout and settle actual usage afterward. Account balances and organization resource groups add separate limits. Confirm the payer before testing.

[ikaijua]

Configuration and extensions

To bill an organization, select it or a resource group under Settings → Application → Billing. Created Spaces, dashboards, and repositories still belong to the personal account.

Self-hosting Chat UI requires OAuth, token scopes, configured models, and the ML_ASSISTANT_MODE build flag. A local UI still launches real HF cloud work.

[ikaijua]

Permissions, data, and limitations

Account permissions determine model/data access, job launches, and repository writes. Gated resources require prior approval as well as token scopes.

Inspect data leakage, splits, and evaluation scripts. Completion does not establish paper reproduction; this remains an experimental mode.

[ikaijua]

Frequently asked questions

Why can’t compute launch? The initial budget is $0; credits and tool permissions are also required.

Should I install the archived repo? Use HuggingChat, or Chat UI for development.

Does the budget cap all costs? No; chat inference and other products are separate.

[ikaijua]

Official references

• Migration notice
• Current code
• Getting started
• Local development