model-trainer

You can fine-tune an LLM on Hugging Face Jobs without local GPUs by submitting SFT, DPO, or GRPO training scripts to managed infrastructure. This requires Hugging Face Hub authentication and a paid plan to execute the jobs.

Yes, you can convert a fine-tuned Hugging Face model to GGUF format for local deployment. The process supports exporting trained models directly to GGUF, enabling usage with local inference engines like Ollama or LM Studio.

You can monitor LLM training progress on Hugging Face Jobs in real-time by integrating Trackio. It provides visualization of training metrics and progress directly from the managed cloud infrastructure during the fine-tuning process.

No, DPO and GRPO training do not require local GPU setup. You can execute direct preference optimization and group relative policy optimization directly on managed Hugging Face infrastructure by submitting jobs remotely.

Supervised fine-tuning with trl and peft requires dependencies including transformers, accelerate, and sentencepiece. These libraries facilitate model loading, distributed training, and tokenization when running SFT jobs on cloud infrastructure.

You need a paid Hugging Face plan for model training jobs because the infrastructure requires dedicated compute resources. Paid plans unlock the Hugging Face Jobs execution environment needed to run supervised fine-tuning and reward modeling tasks.