Own your AI model and its weights.

Bring your own keys. Your model keys never leave your machine. A While. key adds hosting: we store your datasets, run the hosted judge, train and serve.

Where the weights land

1. 01

   A directory on your disk

   A training run writes a LoRA adapter directory. It holds adapter_config.json, which is what marks it a model and not a dataset. Nothing leaves the machine for it to exist, and nothing has to be served for you to keep it.

2. 02

   A model repo on your account

   wai.hub.push uploads that directory to a Hugging Face repo in your own namespace with your own token, private until you say otherwise. Read it back anywhere with PeftModel.from_pretrained.

3. 03

   An endpoint an app can call

   wai.platform.serve hosts a finished run's adapter and returns the row whose endpoint is an OpenAI-compatible base URL and whose name is the model id to send. unserve is the inverse, and the adapter is still yours after it.

4. 04

   The rows that trained it

   wai.export writes the graded rollouts to a file in the format a trainer reads, TRL or a preference set, with push_to for the same account. The dataset is the part a frontier API never hands back.

import whileai as wai

# the graded rows that trained it, on your disk
wai.export(rows, "train.jsonl", format="trl")

# the adapter a run wrote, to a model repo on your own account
wai.hub.push("out/adapter", "me/my-lora")

# the same weights behind an OpenAI-compatible endpoint
model = wai.platform.serve("refund-agent", run)
model["endpoint"]

Rent or own

You rent a model that never learns your business, or you own one that does.

A rented model is the same model for every customer of the vendor, and the next version of it arrives on the vendor's schedule. The prompt is the only thing you can change, and prompts do not learn. An open model post-trained on your own traffic learns the part of the job that is yours, and the next day's traffic is the next training set. While. does not ask you to pick on faith: the before and after are scored on the same held-out set, per task category, with a 95% interval, so a gain in one kind of request cannot hide a loss in another.

Questions

Who owns the weights after fine-tuning?

You own the model, the data and the weights: the datasets are built from your production traces, the model is an open model post-trained with SFT and RL, and the trained weights are yours to download and serve anywhere. A training run writes a LoRA adapter directory that stays on your disk, written by your compute from your data. From there wai.hub.push sends it to a Hugging Face model repo on your own account with your own token, and wai.platform.serve puts the same weights behind an OpenAI-compatible endpoint. The base model's license governs what you may redistribute, and no part of whileai needs a While. account for the adapter to exist. Bring your own keys. Your model keys never leave your machine. A While. key adds hosting: we store your datasets, run the hosted judge, train and serve.

Can I train my own model on production data?

Yes, and that is the intended path. whileai reads your production traces, simulates the situations the agent fails on, grades every rollout with a program where one can score it and a checked judge where none can, keeps the rows that carry signal, and trains an open model on them. The traces stay on your infrastructure unless you hand them to the hosted platform.

Should I rent a model or own one?

Renting a frontier API is the shorter path to a working agent and it stays the right answer while the harness is still moving. Owning becomes the question once the same failures repeat: a rented model cannot learn them, because its weights are not yours to change. whileai is built for the second case, and it measures the first one honestly, scoring a frontier agent on the same held-out set so the comparison is paired rather than asserted.

Can I use an open model instead of a frontier API?

Yes. whileai trains open models with SFT, DPO and GRPO and proves each gain on a held-out set with a 95% confidence interval, so the choice is a measurement and not a preference. Keep the frontier model where it wins; the held-out score per task category says where it does.

What is BYOK post-training?

Post-training where you supply the model keys and the compute, and the library runs against them from your machine. Modal and Prime Intellect are first-class, and a run never requires a While. account. Bring your own keys. Your model keys never leave your machine. A While. key adds hosting: we store your datasets, run the hosted judge, train and serve.

Start with the measurement

Owning a model is worth nothing without a test it never saw. Build that first, on the model you run today, then decide what to train.