01 · TRAIN
On your examples
Fine-tune a supported open base model with LoRA or a full fine-tune, for supervised (SFT) or preference (DPO) training. Training runs on Together AI's infrastructure; you upload, watch and cancel from Yelno.
Your data, your model, your endpoint.
Fine-tune a supported open model on your own examples, serve it on a dedicated endpoint, and call it with the API key you already use. Training runs on Together AI's infrastructure. Hosting is billed per GPU-hour while the endpoint runs, and it is usually the larger cost.
READ THIS FIRST
Hosting is usually the dominant cost. Training is charged once per job; your model's endpoint is billed per GPU-hour for as long as it runs, prorated per minute, whether or not it receives requests. A stopped endpoint costs nothing, and auto-stop is on by default.
How it works
Train it, host it, call it.
Fine-tuning adapts an open model to your own examples: your tone, your formats, your task. Yelno runs the money and the controls around it, with the same API key, policy and wallet as every other model.
02 · HOST
On your own endpoint
The result is served on a dedicated endpoint for your organization, on the hardware you choose, billed per GPU-hour while it runs. Stop it and it costs nothing.
03 · CALL
With the SDK you use
Call it like any other model, with the OpenAI or Anthropic SDK and your existing API key, under an id of the form ft:<base>:<org>:<id>. It is private to your organization.
Pricing
Training once, hosting by the hour.
Final prices only: what you see is what your wallet is charged. Every amount is held before it is spent, and every charge is on your statement line by line.
TRAINING · CHARGED ONCE PER JOB
A final price per 1M training tokens
- Priced by base model and method: LoRA or full fine-tune, each for SFT or DPO.
- Estimate = training tokens × epochs + validation tokens × evaluations, at that price. It is an estimate until the job finishes.
- Before training starts, a hold of the estimate plus 20% is placed on your wallet. Not enough funds: 402, and nothing starts.
- You are charged for the tokens actually trained, with a per-job minimum charge; the rest of the hold is released.
- A cancelled job is charged for the training completed before it stopped. A job that fails on invalid input or a provider error is not charged.
HOSTING · USUALLY THE LARGER COST
Billed per GPU-hour while running
- Your model runs on a dedicated endpoint. It is billed per GPU-hour, at the hardware's final hourly price per replica, while it is starting or running.
- Prorated per minute and posted every 5 minutes. A stopped endpoint costs nothing and holds nothing.
- Auto-stop: after 15 idle minutes by default, at the maximum hours or the budget you set, or when the wallet runs out.
- Starting an endpoint needs at least 1 hour of its hourly price available in your wallet.
- Requests to your own endpoint carry no per-token charge: the hourly price covers them.
Your data
Where your data goes.
Training needs your data to leave Yelno for the training provider, Together AI. That is why fine-tuning is opt-in, and why it is stated here before anything else about it.
Sent to Together AI
Your training and validation files are sent to Together AI, and training runs on Together AI's infrastructure. The console says so before you start a job.
Deleted after the job
Datasets are deleted from Together AI after the job finishes, unless you choose to keep them there. Yelno's own copy follows your organization's file retention.
Off until you turn it on
Fine-tuning is off by default for every organization. An owner or security admin turns it on with the allow_fine_tuning policy, and the change is recorded in the audit log.
Your data controls apply
An organization that requires zero data retention is refused, with the reason, unless the training provider's terms satisfy that requirement.
Private to your organization
A fine-tuned model belongs to your organization. It never appears in the public catalog, and no other organization can call it.
See the subprocessor list and the security page for how the rest of the platform handles content.
Step by step
From dataset to endpoint.
The console walks through each step; every step is also an API route.
- 01
Turn it on
An owner or security admin enables fine-tuning in the organization policy (allow_fine_tuning). It is off by default; turning it on needs a recent sign-in, follows your approval policies, and is audited.
- 02
Upload a dataset
Upload a JSON Lines file of chat, completion or preference examples, with purpose fine-tune, up to 100 MB. Add a validation file if you want loss measured during training.
- 03
Read the validation report
Every file is checked on upload: roles and turn order, tool calls, token counts, duplicates and empty lines. Errors come back with line numbers.
- 04
Create the job
Pick a base model, the method (LoRA or full) and the objective (SFT or DPO). You see the estimate and the wallet hold before anything starts.
- 05
Watch it train
Follow the job's state, events, training and validation loss, and checkpoints. Cancel at any time.
- 06
Deploy an endpoint
Choose hardware at its final hourly price and set auto-stop: idle minutes, maximum hours and a budget.
- 07
Call your model
Use the OpenAI or Anthropic SDK with your usual API key and the model id ft:<base>:<org>:<id>. No per-token charge is added.
- 08
Stop it
Stop the endpoint when you are done, or let auto-stop do it. A stopped endpoint costs nothing, and you can start it again later.
QUESTIONS
What teams ask first.
The guide covers datasets, hyperparameters and endpoints in detail; the API reference lists every route.
- Where is my training data sent?
- To Together AI, the training provider; training runs on Together AI's infrastructure. Uploaded datasets are deleted from Together AI after the job finishes unless you choose to keep them there, and Yelno's own copy follows your organization's file retention.
- Is fine-tuning on by default?
- No. It is off for every organization until an owner or security admin turns on the allow_fine_tuning policy. The change is audited, and every fine-tuning route checks it.
- What does a training job cost?
- The base model's final price per 1M training tokens for the method you choose, times the tokens actually trained. Before the job starts you see an estimate (training tokens × epochs, plus validation tokens × evaluations) and a hold of the estimate plus a buffer is placed on your wallet. You are charged for the tokens actually trained, with a per-job minimum charge shown alongside the price, and the rest of the hold is released.
- How is hosting billed?
- Per GPU-hour while the endpoint is starting or running, at the hardware's final hourly price per replica. It is prorated per minute and posted every 5 minutes. A stopped endpoint costs nothing and holds nothing. Hosting is usually the largest part of the cost, which is why auto-stop is on by default.
- When does an endpoint stop by itself?
- When it has been idle for its idle limit (15 minutes by default), reaches its maximum hours or its budget, or when your wallet can no longer pay for it. Each stop is recorded with its reason and sent as a fine_tuning.endpoint.stopped webhook.
- Do requests to my fine-tuned model cost extra?
- No per-token charge is added: the endpoint's hourly price covers the requests it serves. If no endpoint is running, the API answers 409 model_not_deployed instead of starting one for you.
- Who can call my fine-tuned model?
- Only your organization, with its own API keys. The model never appears in the public catalog, and another organization cannot call it.
- Can an organization that requires zero data retention use it?
- Only if the training provider's terms satisfy that requirement. Otherwise creating a job is refused with the reason, before any data leaves the platform.
- What if a job fails or I cancel it?
- A cancelled job is charged for the training the provider completed before it stopped. A failed job is charged only if the provider bills for it; a failure caused by invalid input or a provider error is not charged. The rest of the hold is released. If the provider's final figures are missing, the job is reviewed rather than settled on a guess.
- Is there a quality or uptime guarantee?
- No. Results depend on your data and settings, so no accuracy or quality is promised, and no availability or SLA is offered for endpoints. Every estimate is labelled as an estimate.
No accuracy, quality, availability or SLA is promised for fine-tuned models or their endpoints. Prices are final customer prices, published with an effective time; estimates are labelled as estimates, and the charge for a job or an endpoint is on its statement line.