Most ML engineers spend more time debugging training scripts than actually training models.

Writing training code by hand is the bottleneck nobody talks about

Setting up a fine-tuning run means choosing the right GPU tier, configuring LoRA versus full fine-tuning, wiring up authentication, and wrestling with a dozen other decisions before a single token gets trained. One wrong flag and the job silently fails at hour three on a paid GPU.

Claude now submits the job, watches it run, and ships the model

We Got Claude to Fine gives Claude a packaged skill called hf-llm-trainer that contains the full domain knowledge of a training run: hardware selection, script configuration, Trackio monitoring, and Hub deployment. You tell it something like “fine-tune Qwen3-0.6B on this dataset,” and Hugging Face Skills handles hardware selection, submits the job to Hugging Face cloud GPUs, reports the job ID and estimated cost, and delivers a finished model to your Hub profile when the run completes.

ML teams shipping fast are the first to feel this shift

• ML engineers who burn hours configuring training runs instead of designing experiments get those hours back immediately.
• Applied AI researchers who need to test supervised fine-tuning, DPO, and RLVR pipelines on the same dataset without rewriting scripts each time.
• Startup technical leads who need production-grade fine-tuned models but cannot justify a dedicated MLOps hire.

The skill supports models from 0.5B to 70B parameters and includes GGUF conversion for local deployment after training.

Agentic coding tools just crossed into ML infrastructure territory

Claude Code, Codex, and Gemini CLI are already supported, with Cursor, Windsurf, and Continue integrations coming, which means the same skill distribution model that took over frontend development is now pointed directly at ML workflows. Teams that relied on internal training tooling as a competitive advantage are about to face a much shorter gap to close.

What you can actually run with this today

• Submit a fine-tuning job to cloud GPUs with one natural language prompt.
• Run supervised fine-tuning, DPO, or RLVR on the same base model.
• Convert any finished checkpoint to GGUF for local inference.
• Chain multi-stage training pipelines inside a single Claude Code session.

Requires a Hugging Face Pro or Team plan for Jobs access; the skill itself is free via the marketplace.

The tool only works inside supported coding agents and requires a write-access Hugging Face token, so teams not already using Claude Code or a compatible CLI will need to onboard before they can run anything.

If you want something closer to a GUI, Hugging Face AutoTrain handles fine-tuning through a web interface. For pure script-based control without an agent layer, Axolotl remains the go-to open source trainer.

AI coding agents are quietly absorbing the MLOps stack

This is no longer about autocomplete or generating boilerplate. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.