Job outputs

When an alallo job's output is ready

When an alallo training job finishes successfully, its status becomes "succeeded", its progress shows 100%, and the job detail in the dashboard shows "Training finished. The fine-tuned model is ready." Only succeeded jobs have an output. Cancelled and failed jobs do not produce a fine-tuned model. [TBD: whether customers can get the last saved checkpoint of a cancelled or failed job.]

What an alallo LoRA, QLoRA or DPO job produces

A job using LoRA (Supervised fine-tuning, full precision 16-bit), QLoRA (Supervised fine-tuning, 4-bit) or DPO (Direct preference optimisation, LoRA) produces a LoRA adapter: a small set of extra weights, together with the tokenizer files of the base model. The adapter is not a complete model. To use the adapter, load the original base model (for example Qwen/Qwen2.5-7B-Instruct from Hugging Face) and apply the adapter on top. The adapter is saved in the standard Hugging Face PEFT format (an adapter configuration file and an adapter weights file). [TBD: confirm the exact file list and that weights are in safetensors format.]

What an alallo full fine-tuning job produces

A job using Supervised fine-tuning (full weights) produces a complete fine-tuned model: all model weights and the tokenizer files, in the standard Hugging Face Transformers format. A full model is as large as the base model and can be used on its own, without the base model.

How long alallo keeps your fine-tuned model

alallo keeps the fine-tuned model of a succeeded job for 30 days after the job finishes, at no extra cost: storage is not charged in alallo tokens. After 30 days, alallo deletes the model. Download the model, or copy it to your own storage, within those 30 days if you want to keep it. A free Hugging Face account, for example, includes private model storage. [TBD: whether alallo warns before deleting a model, and whether the 30 days can be extended.]

What alallo deletes when an alallo job ends

While a job runs, alallo saves training progress (checkpoints) so that the job can continue if it has to be moved. These checkpoints are only needed while the job runs. alallo deletes them when the job ends, and keeps only the fine-tuned model of a succeeded job. Cancelled and failed jobs keep no model.

Downloading an alallo model

To download the fine-tuned model of a succeeded alallo job, open the job in the Jobs tab and click "Download model". alallo lists the model's files with their sizes; click each file to download it. The download links are created when you click "Download model" and stop working after a short while; click "Download model" again (or reload the page) to get fresh links. The job also shows the date until which alallo keeps the model.

The "Your models" tab, next to "New job", shows every fine-tuned model in your account that can still be downloaded, with the date until which it is kept and a Download button. Every user in your alallo account can see and download the account's models. Only your account can reach its models: no other alallo customer can list or download them. [TBD: whether alallo can copy the model to the customer's Hugging Face account.]

Using an alallo LoRA adapter

A LoRA adapter from alallo can be used with common open-source tools that support Hugging Face PEFT adapters, for example the Hugging Face transformers and peft Python libraries, and inference servers that support LoRA adapters. Load the same base model that the job fine-tuned, then load the adapter. Many tools can also merge the adapter into the base model to create a single standalone model. [TBD: whether alallo offers a merged-model output or hosted inference.]

Prompt format for an alallo fine-tuned model

An alallo fine-tuned model learns exactly the text format your training data used. alallo does not add any chat formatting during training. When you use the fine-tuned model, send prompts in the same format as your text template, up to the point where the answer starts. For example, if you trained with the template ### Instruction: {instruction} ### Response: {output}, send ### Instruction: <your question> ### Response: and let the model complete the answer.

Licence of an alallo fine-tuned model

A model fine-tuned with alallo is based on an open model and remains subject to the base model's licence (for example the Qwen2.5 or Phi-3.5 licence on Hugging Face). [TBD: alallo's terms on ownership of fine-tuned models and adapters.]