Do I need machine learning experience?

No. You need your agent’s tools and system prompt, and a way to say what a good answer looks like. Zero Proof does the simulation, the scoring and the dataset.

What is the difference between RL and SFT here?

Supervised fine-tuning (SFT) shows the model good conversations to copy. Reinforcement learning (RL) gives the model a score for each attempt so it can find better ones: GRPO samples during training and rewards the right first action, DPO trains offline on each prompt's pass against its fail. All three come from the same simulations and are one Train button or one zps.train call, so you can start with SFT and move to RL without new data.

Do I need GPUs to start?

No. Simulation, scoring, training and hosting run on our side. If you already have a training setup, pull the dataset and use that instead.

How do I know training helped?

Every dataset ships with a test set that was never trained on. Run it before and after. If the pass rate did not move, the training run did nothing, and you know that before it reaches customers.

You do. The weights are yours to download, and the endpoint is yours to keep or move.

Which models and tools does it work with?

Any OpenAI-compatible endpoint, vLLM, Ollama, or our hosted model. Production comes in over OpenTelemetry. Datasets go out to Hugging Face or a JSONL file.

ZZero Proof Labs

Platform Docs Research

Post-training platform

Train and host your own AI models

You are responsible for your company’s AI. Zero Proof turns what your agent does in production into training data, trains a model that is yours with reinforcement learning (RL) or supervised fine-tuning (SFT), hosts it, and shows you the before and after.

Get started Datasets on Hugging Face

$ pip install zeroproof

train.pyzeroproof 0.27

import zeroproof.simulations as zps
# 1. generate conversations from your agent's tools and prompt
data = zps.simulate(tools=my_tools, system_prompt=my_prompt, mode="rl")
# 2. score every conversation with your grader
scored = data.grade(judge=my_judge)
scored.pass_at            # how often it passes
# 3. keep the useful ones and publish the dataset
rows, report = zps.optimize(scored, mode="rl")
entry = zps.push_rows(rows, "support-rl-v1", gate=True, mode="rl")
# 4. train; the run reports its loss curve to the platform
run = zps.training_run("support-rl-v1", dataset=entry["datasetId"], base_model="Qwen/Qwen3-4B")

Works withOpenTelemetry Any OpenAI-compatible model vLLM Ollama Hugging Face Prime Intellect environments

Platform

From production to a model of your own, in six steps

01Connect

Connect production

Set three environment variables in your app. Every conversation your agent has shows up here with a score.

OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=https://api.zeroproofai.com/v1/traces

02Diagnose

See what fails

A list of the requests your agent gets wrong and how often, grouped by what the customer asked for.

zps.trace_report(rows)

03Generate

Generate training data

We simulate your agent against those situations, with made-up customers and tools that fail on purpose, and score every conversation. No annotators.

zps.simulate(tools=my_tools, system_prompt=my_prompt)

04Train

Train a model of your own

One call or one button. Supervised fine-tuning (SFT) on the good conversations, GRPO (on-policy RL against a verifiable reward) or DPO (offline, each prompt's pass against its fail). Loss, reward and pass@1 before and after on the run page. Or pull the dataset into your own training setup.

run = zps.train("ds_…", method="grpo", holdout="ds_…")

05Host

Run it on our endpoint

Your model is served behind an OpenAI-compatible URL. Swap it into your app by changing one setting. The weights stay yours.

OpenAI(base_url="https://…modal.run/v1")

06Verify

Prove it worked

Every dataset ships with a test set the model never saw. Run it before and after training. The number moves or it does not.

zps.delta_report(before, after, target="pass_at_1")

5.0% to 30.0%: pass rate of an open 8B model on a public customer-support test, after one training run on our data
$15: what that training data cost to generate. No annotators, no human-written examples.
75.3%: of refund decisions correct from a 1B model that costs 18 cents per million tokens to run
3 settings: in your app to connect production. No code change.

Results

Two public results

more tasks passed after one training run on a $15 dataset

open 8B model
support test

No training data5.0%
1,057 examples written by experts17.5%
1,057 examples from our simulations, $1530.0%

Same number of examples as the expert set, none written by a person. Method

1/100th

the cost of frontier models, 5 to 13 points behind them on refund checks

1B model
right answers vs cost

Trained on simulated marketplace conversations. The model, data and test set are public. Method

You own the company’s AI

You need a model that is yours, gets better every month, and comes with a number that proves it did.

Get started

You ship an agent to customers

You need to know before customers do when a prompt change or a model upgrade made things worse.

Connect production

You train models

You need scored data with a clean test set, not a pile of logs. Pull a dataset and go.

Datasets on Hugging Face

Datasets

Ready-to-train datasets

Scored datasets for customer-support and coding agents. One call optimizes a set for reinforcement learning (RL) or supervised fine-tuning (SFT): it keeps the examples a model can still learn from and drops the rest. Each ships with a test set, and one call publishes it to Hugging Face or imports one from there.

import zeroproof.simulations as zps rows = zps.pull("ds_…") # no key needed rows, report = zps.optimize(rows, mode="rl") # keep what a model can learn from

Datasets on Hugging Face

Research

All posts

GuideSep 14, 2026

Character training with Zero Proof

A constitution, a judge that reads it, length-matched pairs, and a before-and-after number.

DocsSep 13, 2026

How training data gets made

How one training example gets made, shown on a real airline agent.

PostSep 9, 2026

Simulated data pipeline for fine-tuning

A $15 simulated training set beat expert-written examples on a public support test.

PostJul 13, 2026

Safety judges for agentic commerce

A 1B refund checker trained inside a simulated marketplace.

Questions

Do I need machine learning experience?: No. You need your agent’s tools and system prompt, and a way to say what a good answer looks like. Zero Proof does the simulation, the scoring and the dataset.
What is the difference between RL and SFT here?: Supervised fine-tuning (SFT) shows the model good conversations to copy. Reinforcement learning (RL) gives the model a score for each attempt so it can find better ones: GRPO samples during training and rewards the right first action, DPO trains offline on each prompt's pass against its fail. All three come from the same simulations and are one Train button or one zps.train call, so you can start with SFT and move to RL without new data.
Do I need GPUs to start?: No. Simulation, scoring, training and hosting run on our side. If you already have a training setup, pull the dataset and use that instead.
How do I know training helped?: Every dataset ships with a test set that was never trained on. Run it before and after. If the pass rate did not move, the training run did nothing, and you know that before it reaches customers.
Who owns the model?: You do. The weights are yours to download, and the endpoint is yours to keep or move.
Which models and tools does it work with?: Any OpenAI-compatible endpoint, vLLM, Ollama, or our hosted model. Production comes in over OpenTelemetry. Datasets go out to Hugging Face or a JSONL file.

Start

Get started in three commands

You get an account and an API key from the terminal. Free to start.

Get started Talk to us

1Installpip install zeroproof
2Sign upzeroproof signup --email you@company.com
3First runzps.simulate(spec="specs/github", budget=20)