Hugging Face AutoTrain Is Deprecated. Here Is What to Use Instead.

TL;DR

Hugging Face AutoTrain has been deprecated. The official documentation now redirects users to Axolotl, TRL, and the Transformers Trainer. The product is no longer being maintained, bugs are not being fixed, and support is non-existent. For engineers who built fine-tuning workflows around AutoTrain, or who are just discovering this and need to know what to do next, this post covers your actual options: what each alternative does, who it is built for, and how to choose based on your situation.

If you have been using Hugging Face AutoTrain and something stopped working recently, you are not imagining things and you are not doing anything wrong.

AutoTrain is deprecated. The official documentation page now reads: "This project is no longer maintained. No new features will be added and bugs will not be fixed."

Hugging Face's own recommendation is to move to Axolotl, TRL, or the Transformers Trainer.

That is a significant change for anyone who relied on AutoTrain's no-code workflow to fine-tune domain-specific models without managing a training stack. And the Reddit threads make clear that a lot of engineers are stuck right now. They know AutoTrain is broken. They do not know what to replace it with. And the options they are finding require significantly more setup than uploading a CSV and clicking submit.

This post is the clear answer to that question.

Why AutoTrain Is Gone

AutoTrain launched as a genuinely useful no-code fine-tuning tool. Upload your data, pick a base model, get a fine-tuned model back. For small teams and non-ML engineers it was a real unlock.

But the product was never deeply invested in. The underlying framework accumulated technical debt. When the ML ecosystem moved fast, AutoTrain did not keep up. Dependencies drifted. Bugs accumulated. Support tickets piled up without answers.

The GitHub repository for AutoTrain Advanced has open issues. Many of them are answered not with fixes but with "please open a separate issue" or no response at all. The Discord reports the same pattern. Users who reached out to Hugging Face support were told to file GitHub issues. Those issues sat open.

Hugging Face made the right call by officially deprecating it rather than leaving engineers guessing why it was broken. But that does not make the transition easier for people who need a working fine-tuning workflow right now.

What Hugging Face Recommends Instead

The official recommendation from Hugging Face is three tools: Axolotl, TRL, and the Transformers Trainer.

Here is what each one actually is and who it is for.

TRL (Transformer Reinforcement Learning) is a Hugging Face library for fine-tuning with SFT, DPO, GRPO, and reward modeling. It is code-first. You write Python scripts using TRL's trainer classes. If you are comfortable writing training scripts and want to stay close to the Hugging Face ecosystem, TRL is the most direct migration path from AutoTrain. But it is not no-code. You need to understand enough about training to write and debug a Python script.

Transformers Trainer is the base training class inside the Transformers library. It handles the training loop and gives you a lot of control over the process. It is more verbose than TRL and requires more setup, but it is also the most flexible option and the one with the deepest documentation. If your use case has specific requirements that TRL does not cover, the Transformers Trainer is the escape hatch.

Axolotl is the recommended option for most engineers coming off AutoTrain who want a real fine-tuning workflow without writing Python training scripts from scratch. You define your training job in a YAML config file. Axolotl handles the framework, the distributed training configuration, checkpointing, and logging. It supports QLoRA, LoRA, full fine-tuning, SFT, DPO, and more. The config format is readable, well-documented, and widely used in the open-source fine-tuning community.

The Real Problem With Moving to These Tools Directly

Here is the thing the official recommendation does not address.

TRL and Axolotl are excellent tools. But they require a GPU environment to run on. And setting up a GPU environment from scratch is the exact problem AutoTrain solved for people.

You need a machine with a compatible GPU. You need CUDA installed and matched to your driver version. You need your Python dependencies pinned correctly. You need to handle checkpointing so a failed job does not lose your progress. And when something breaks, which it will on a first setup, you need to debug it.

For engineers with ML infrastructure experience, this is a few hours of work. For the engineers who chose AutoTrain specifically because they did not want to manage this layer, it is a significant new requirement.

The Reddit threads about AutoTrain's deprecation are full of this exact frustration. People who were successfully running dozens of fine-tuning jobs on AutoTrain, refreshing models every quarter as their datasets updated, are now stuck at the infrastructure layer before they can run a single job.

The Path Forward Depending on Where You Are

If you are comfortable writing Python training scripts: TRL is your migration path. The SFTTrainer class covers most supervised fine-tuning use cases and the documentation is solid. You will need a GPU environment but if you are already writing Python you can manage this.

If you want a config-based workflow similar to AutoTrain but with more control: Axolotl is the answer. You write a YAML config that describes your training job. Axolotl handles the training framework underneath. It is closer to AutoTrain in workflow than TRL because you are not writing Python. But you still need a GPU environment to run it on.

If you want the simplicity of AutoTrain's submit-and-wait workflow without the broken product: A purpose-built fine-tuning platform is the closest replacement. Together AI offers managed LoRA fine-tuning with minimal setup. Heulistic is built around Axolotl with managed infrastructure underneath, which means you get the full Axolotl config workflow without having to provision a GPU instance or install CUDA. You upload your config, see a cost estimate, and submit. The infrastructure is handled.

Who This Is Hardest For

The engineers hit hardest by this are the ones the Reddit threads describe. Small teams, often a team of one, running many fine-tuning jobs on a regular refresh cadence. Classification models, NER, text tasks across dozens of domain-specific datasets that update quarterly. AutoTrain made this manageable without ML infrastructure expertise. Nothing in Hugging Face's official recommendation replaces that workflow cleanly.

If this describes your situation, the honest answer is that you have two realistic paths.

You invest the time to build a proper fine-tuning workflow. That means learning enough about Axolotl or TRL to write configs or scripts, and setting up an environment to run them on. The upfront cost is real but the workflow you build is more robust than AutoTrain ever was and gives you control AutoTrain never did.

Or you find a platform that removes the infrastructure layer so the workflow is the config, not the environment setup. That is what Heulistic is built for. If you were using AutoTrain to run Axolotl-compatible fine-tuning tasks, Heulistic is the closest replacement to what you had in terms of simplicity while being an actual upgrade in terms of control and reliability.

The One Thing Worth Knowing

AutoTrain's deprecation is not just a product decision. It is a signal about where the fine-tuning ecosystem is heading.

The tools that are being actively invested in are code-first and config-first. Axolotl, TRL, Unsloth. The no-code wrapper layer is not where the ecosystem is putting its energy. If you are building a fine-tuning workflow that needs to last, building it on one of those foundations rather than on a managed abstraction that can be deprecated is the more durable choice.

Axolotl is the foundation that matters. Whether you run it yourself or on a platform that handles the infrastructure underneath, your config is the portable artifact that describes your workflow. That config will run on Axolotl anywhere. That is a more durable foundation than any specific platform.

You can get started with Heulistic at heulistic.com.