Toloka review: a training data company for agent fine-tuning on a budget
Answer in brief: Toloka is our runner-up in two categories: enterprise fine-tuning and best budget option. Its fine-tuning service builds expert-corrected data, RL Gym environments and a held-out eval, then post-trains a small open model to match a frontier API on your narrow task and hands back the weights. Its self-service platform has no minimums or contracts. It charges per item for expert work and per GPU-hour for fine-tuning, plus a platform percentage, and shows an estimate before you launch. This review is independent research from public sources.
Best-for verdict: runner-up, enterprise fine-tuning. Runner-up, budget. Visit Toloka →
What Toloka does for training data buyers
- Enterprise agent fine-tuning: SFT first, then gisting, with RL only where the eval shows it pays off. Training runs on Nebius compute, and you get either an OpenAI-compatible endpoint or the weights.
- RL environments and agent data: virtual companies, RL gyms with MCP replicas and computer-use testbeds, and agent trajectory demonstrations.
- Evaluation and red-teaming: including prompt-injection and policy-compliance testing.
- Coding data, LLM/VLM datasets, physical-AI and egocentric video data, plus off-the-shelf datasets and Toloka Arena.
- Platform: describe a data goal, and an agent builds the collection, annotation and QA pipeline. Experts work through Mindrift.
Who it fits
Toloka fits enterprises paying too much for frontier APIs on repetitive agent tasks, agent builders who need realistic environments, and smaller teams that want self-serve data without a contract.
Why it stands out
- A named enterprise case. Shopify routed de-identified Sidekick conversations to Toloka experts to build SFT golden sets. The segmentation judge's average class accuracy rose from 75% to 89%.
- Cost proof. In Toloka's own Mindrift CV-parsing deployment, inference cost fell from $10 to $30 per 1,000 documents on a frontier API to about $0.80 on a fine-tuned model, with F1 of 0.94 versus 0.93 (vendor-reported).
- Compliance. ISO 27001, ISO 27701, SOC 2 Type II, GDPR, CCPA and HIPAA are listed on its fine-tuning page.
- Low-friction buying. Prepaid self-service, with charges captured on delivery and unused reserves released.
The catches
- No universal rate card. Toloka shows rates in your account or at checkout, not publicly.
- Anonymous testimonials. Homepage quotes are unattributed, so ask for references.
- Smaller scale than the top-tier vendors by reported revenue.
- Final purchases. Self-service purchases are generally non-refundable under its agreement.
Pricing
Self-service: per item for expert work, per GPU-hour for fine-tuning, plus a platform percentage, with no minimums or contracts. Managed projects are quoted.
Best alternatives to Toloka
- Scale AI, our enterprise fine-tuning winner.
- Prolific, our budget winner, with a published platform fee.
- Invisible Technologies, for fast expert mobilization.
Toloka FAQ
Does Toloka have self-service pricing?
Yes. It has no minimums or contracts. Expert work is billed per item and fine-tuning per GPU-hour, plus a platform percentage.
Do I get the model weights?
On Toloka's full-service option, yes. It hands back the weights of the fine-tuned open model.
What is Mindrift?
Mindrift is Toloka's expert network.
Sources
- Toloka enterprise agent fine-tuning
- Toloka for AI engineers
- Toloka platform
- Toloka Self-Service Agreement
- Shopify and Toloka case
← Back to the ranked list of the best training data companies