Skip to content
tokz labs
Start free

Replace an LLM call with a small model you own

Fine Tune trains your own copy of OpenDecision-Small on your text and one question, then scores it against OpenDecision-Small on the same rows.

Start with $5 of creditRead the Fine Tune docs
$2.00
only if it beats OpenDecision-Small
3 to 10 min
a job, in our tests
31 MB
the demo spam filter
$0
a call on your machine
Spam filter, macro F1 out of 100
  • spam-filter, tuned99.2
  • OpenDecision-Small48.3
Both models scored on the same 240 held-out messages, labeled by AI. About 42 ms a message on one CPU thread.

Watch the spam filter run

How does Fine Tune work?

In the dashboard, open Fine Tune and choose New fine tune. We email you when it finishes.

  1. 01

    Say what to decide

    Name the model and write one question, like "Which team should handle this ticket?"

  2. 02

    Add your file

    CSV, TSV, JSON, JSONL, Excel, Parquet or plain text. Up to 5 files, 20 MB.

  3. 03

    Answer 20 rows you know

    For a text-based file, the form shows its rows one at a time. These count for more than typed rows.

  4. 04

    Call it by name

    On our servers, or free on your own machine.

When does fine tuning help?

It paid on a few labels, with real text, when OpenDecision-Small was weak. It did not pay with many labels, or when the labels did not follow from the text.

Fine Tune on public datasets
Emotion, 6 labelsSmall 47.2tuned 59.8Better
Financial sentimentSmall 60.8tuned 71.8Better
Spanish reviews, 3 labelsSmall 57.1tuned 65.8Better
Arabic tweets, 2 labels, AI labelsSmall 40.8tuned 70.7Better
Japanese reviews, 3 labelsSmall 57.5tuned 62.9Not clear
20 Newsgroups, long textSmall 59.0tuned 59.1Not clear
28 emotions, multi-labelSmall 29.1tuned 30.9Not clear
Tickets, messy Excel file, 10 queuesSmall 21.7tuned 15.6Not clear
Banking, 77 labels, true labelsSmall 70.3tuned 69.1Not clear
Banking, 77 labels, AI labelsSmall 70.9tuned 66.2Worse
Macro F1 out of 100 on 200 rows the model never trained on, one run each, measured on 2026-11-03. Seeds move these numbers. None of this is customer data, so test on your own.

What does Fine Tune cost?

Starting any fine tune needs $2.00 in your balance.

Fine Tune prices
Fine tune with a file$2.00Only if it beats OpenDecision-Small on rows you labeled.
Fine tune with no file$0No charge.
Hosted calls$0.025Per 1M billed tokens, at least 256 tokens a request, the same as OpenDecision-Small.
Local calls and downloads$0Free.

See all prices

How do I call my model?

By name, through the Tokz API or on your machine. The result has the same shape, and a local call is free.

TypeScript · @tokz/sdk
import { Tokz } from "@tokz/sdk";

const tokz = new Tokz({ apiKey: process.env.TOKZ_API_KEY! });

const hosted = await tokz.decide("support-triage", "I was charged twice.");
const local = await tokz.decideLocal("support-triage", "I was charged twice.");

Python has the same two calls: decide and decide_local.

Fine Tune questions

What is Fine Tune?

A Tokz Labs service that turns an LLM call that picks from a short list into a small model you own. After training, a call needs only the model's name and the text.

Which tasks fit?

Any LLM call that picks from a short list: spam, sentiment, routing tickets to a few queues, flagging risky requests.

What if my model does not win?

You pay nothing. Looks better, No clear difference and Worse are not charged, and each leaves a model you can use or delete. A job that fails or that you cancel is not charged either.

How many rows do I need?

For a real score, 300 labeled rows or 800 unlabeled rows. Rows you know the answer to can help more than anything else you add.

Who sees my text?

Your text goes to RunInfra, Freestyle and Modal to build the model, and it trains only your model. Your files never reach our disk, our database or our logs.