Data Solutions

Enterprise

Physical AI

Platform

Resource Hub

Company

Arena

Talk to us

Toloka Train

Lower the cost of every AI request

Stop overpaying for expensive AI requests. Upload your data to compress your prompts or fine-tune a model for your specific task. Use Toloka Train directly, or work with our team if you’d prefer a managed service

Trusted by Leading AI Teams

Two separate tools, with different inputs and different outputs.
Choose the one that fits your use case.

Fine-tuning

Prompt Gisting

Pick LoRA fine-tuning.
Trains a small adapter on top of a frozen base model (Qwen3, 4B–235B)

Toloka Train Fine-tuning

Replace an expensive frontier model with a smaller open model that meets the quality bar your task actually needs.

Get started with Fine-tuning

You’re paying frontier-model prices for a narrow, repeatable task, or spending more than you need to on an open model that’s already in production. Fine-tuning trains a smaller open model to handle the same task at the quality you need without the cost scaling with every token. A lightweight adapter trains on top of a frozen base model such as Qwen3, from 4B to 235B, so you can improve task performance without paying for a full fine-tune.

Match or beat the quality you have now, at open-model serving cost
Train a LoRA adapter on a frozen base — no full retrain
Turn variable per-token spend into fixed, predictable compute

How we used it ourselves

A CV parsing pipeline for Mindrift, Toloka's own contributor network.

12×–37×

cheaper inference vs. the frontier API on the same task

$10–$30 per 1k CVs → ~$0.80 per 1k CVs

0.94 F1

on the human-labeled golden set

up from a 0.85 frontier-model baseline; latest frontier model tested, 0.93

A few days

end-to-end, from data curation to deployed endpoint

>100% price-performance improvement over standard frontier models

Cost becomes effectively fixed compute instead of variable per-token spend — predictable at scale.

These results come from Toloka’s own production work on CV parsing for the Mindrift contributor network. The approach used synthetic labels from a frontier-model ensemble, a human-labeled hold-out set, and a small open-weight model. The results reflect the performance of that specific workload and shouldn’t be treated as a guarantee for every task.

Fine-tune your model

Two separate tools, with different inputs and different outputs.
Choose the one that fits your use case.

Fine-tuning

Prompt Gisting

Pick LoRA fine-tuning.
Trains a small adapter on top of a frozen base model (Qwen3, 4B–235B)

Toloka Train Fine-tuning

Replace an expensive frontier model with a smaller open model that meets the quality bar your task actually needs.

Get started with Fine-tuning

Match or beat the quality you have now, at open-model serving cost
Train a LoRA adapter on a frozen base — no full retrain
Turn variable per-token spend into fixed, predictable compute

How we used it ourselves

A CV parsing pipeline for Mindrift, Toloka's own contributor network.

12×–37×

cheaper inference vs. the frontier API on the same task

$10–$30 per 1k CVs → ~$0.80 per 1k CVs

0.94 F1

on the human-labeled golden set

up from a 0.85 frontier-model baseline; latest frontier model tested, 0.93

A few days

end-to-end, from data curation to deployed endpoint

>100% price-performance improvement over standard frontier models

Cost becomes effectively fixed compute instead of variable per-token spend — predictable at scale.

Fine-tune your model

How much does a run cost?

You're only charged for the GPU-seconds you use. Not for the estimate, and never for a run we rejected.

Held, then reconciled

Based on your parameters, we estimate how long the process will take and place the estimated amount on hold as a safety buffer. Once the process finishes, you’re charged only for the GPU-seconds used, and any unused balance is automatically returned to your account.

Capped, so you never overpay

Runs are strictly capped at six hours, so you never overpay. Every finished run reports the exact GPU-seconds consumed.

Rejected before charged

If there’s an issue with your data, we’ll reject the submission and explain why before you’re charged. You’ll know within seconds, before any cost is incurred.

Toloka Train in production

Explore the fine-tuning case behind the numbers on this page and why strong launch performance doesn’t always last.

Two ways to cut your inference bill — now self-serve on Toloka

How Shopify and Toloka built a ground-truth flywheel to keep skill agents accurate at scale

Overperforming Frontier Models by 100+%: Building a Production CV Parser with Shopify's Tangle

Use your own data and see how the results compare on your workload.

Training is a one-time cost. What changes permanently is your inference bill.

Fine-tune your model

Compress your prompt

Want it fully managed?

Hand us the workflow and get back a production-ready model, built on expert-corrected data, RL Gym environments, and a defensible eval.

Talk to an expert