FAQ

What you get: the questions people ask before their first run, each answered against the released package rather than against the roadmap. Who it is for: AI researchers, ML engineers and applied-AI developers. Needs: nothing, including for the program below. Who is it for? AI researchers, ML engineers and applied-AI developers, and the data teams that own the training data. You do not need a research team to run a training loop. What is wai? wai is While’s whale and the alias of the whileai SDK: import whileai as wai. The company is While, the package on PyPI is whileai, and the command is wai. See install for the line that brings all three. Do I own the model and the data? 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. 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. On hosted GPUs wai.platform.train(dataset_id, method="sft") returns a run whose run.adapter names where the trained weights landed, and train and prove runs the same step on your own GPU, writing the adapter to a path you pick. Do I need a key? Not to start. simulator=False writes the situations from templates instead of from a model, seeded_agent stands in for an agent, and a lambda can be the judge. That is the whole loop with no account and no network:

import whileai as wai


@wai.tool  # the function is the tool; its signature is the schema
def get_order(order_id: str) -> dict:
    """Look up an order by id."""
    ...


data = wai.simulate(
    wai.seeded_agent([get_order]),  # stand-in agent, fails a labeled share
    tools=[get_order],
    system_prompt="Help customers with orders.",
    simulator=False,  # template situation writer: no model, no network
    mode="rl",
    repeats=4,
    repeat_policy="fixed",
    budget=32,
    seed=0,
)
scored = data.grade(judge=lambda row: {"reward": int(not row["seeded"])})
print(scored.pass_at)

pass@1 0.69 [0.50..0.84] | pass^4 (pass_pow_k) 0.25 [0.00..0.50] | pass@4 1.00 [1.00..1.00] | headroom 0.31 (8 groups, k=4)

The number to read is pass@1 0.69, 95 percent interval 0.50 to 0.84, over 8 task groups at k=4. pass@k is the share of tasks solved within k tries, from Chen et al. [1]. The interval is wide because 8 groups is a small set, which is the point: one run of a small set proves little, and the page that sizes the set is how much to run. A key adds the hosted situation writer, the hosted judge, dataset storage, training and serving. See install. Mid-training or post-training? Both, though the post-training methods are the ones this package runs. Mid-training continues training on your domain data so the model knows your world. Post-training with SFT, DPO [2] or GRPO [3] teaches it how to behave. Most teams start with post-training. wai.methods lists what ships. Do I have to change my agent? No. The SDK reads what the agent does, from its traces or from its tools and policy. The agent keeps running where it runs today. See connect your agent. What if my agent is not in production yet? Describe the behavior you want in a sentence. That sentence is the policy. The SDK drafts the tools such an agent would have and runs the conversations against them, so there is training data before there is traffic. How do I know it got better? Every run ends with pass@1 per task category, before and after, on a held-out set with a 95 percent confidence interval. If the interval includes zero, nothing ships.

Default	Value	Where it comes from
`CI_LEVEL`	0.95	`1 - ALPHA`, so the interval and the verdict agree: a delta whose interval excludes zero at `CI_LEVEL` is the one a test at `ALPHA` rejects
`BOOTSTRAP_DRAWS`	2000	resamples behind a percentile interval. Efron and Tibshirani put the floor at 1000; convention above the floor, untested against 1000

Both are in whileai/simulations/defaults.py. Measurement is the Evaluation chapter of the RLHF book [4]. Which models can it train? Open models you own at the end. Training runs SFT, DPO, GRPO or a reward model on hosted GPUs, as a LoRA adapter [5]. Serving is narrower than training: an adapter reaches an endpoint only on a served base, today Qwen/Qwen3-4B and microsoft/phi-4, and wai.platform.train warns when a run will not be servable. See the platform reference. Where do the rollouts come from? From your agent’s own tools and policy. A separate model plays the customer, a mock world answers the tool calls, and every conversation is scored against a written rubric before it is kept. See how it works. Is the judge another LLM? Yes, when you do not pass a program. So it is measured against gold labels, probed with known hacks, versioned by rubric hash, and drawn from a different model family than the policy. See the engine. What happened to zeroproof? That was this package’s name before 0.51, and the company’s name before September 2026. Both are retired. The last zeroproof upload on PyPI installs whileai; every release since ships only as whileai, reads only WHILEAI_* variables and ~/.whileai, and talks to while.ai.

Run the quickstart. It is the program above with your own agent in place of the stand-in.

References

Chen, M. et al. Evaluating Large Language Models Trained on Code. arXiv:2107.03374, 2021. pass@k.
Rafailov, R. et al. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. arXiv:2305.18290, 2023. DPO.
Shao, Z. et al. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300, 2024. GRPO.
Lambert, N. Reinforcement Learning from Human Feedback, Evaluation. rlhfbook.com/c/16-evaluation.
Hu, E. J. et al. LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685, 2021. The adapter the hosted trainer writes.

Get started

Guides

Concepts

Recipes

Reference

Next

References

​Next

​References

Next

References