Start a training run in four steps and follow it while it trains.
New fine-tune is where you start a training run. Open it with the New
fine-tune button in the top bar, with Start fine-tune on
Home, with Fine-tune with these on a
dataset, or with New version on a model.
For the concepts behind it, see the Fine-tune a model guide.
Would you rather not do it step by step? The AI assistant on
Home turns a spreadsheet or text into training data, picks the model, prices the
run and prepares the fine-tune for you to confirm.
Describe the task in your own words. This is marked Coming soon: the note
is for you and your team and does not change training.Model decides where the result goes:
New model starts a model of its own, as v1. Its name is optional: left
empty, it is named after its dataset. If you already have a model with that
name, the form says so and offers to make this its next version instead.
New version of a model adds the next version to one of your models
(for example Lead conversion · v3). It keeps the model’s name, and it can
use different examples or another base model.
Pick a file from My datasets, or switch to Upload new to add one now.Test it on decides how the new version is tested against its base model:
10% or 20% of your examples, held back from training, or Don’t
test. Testing needs at least 50 examples. See
Test results.While you test, Checks and targets lists checks suggested from your
examples. Tick the ones you want and give each a Target: the share of
held-back answers that must pass, in percent. Leave the target empty for a
check you only want measured. The version page shows each target met or
missed, see Checks.
Check
Suggested when
Starting target
Label accuracy
Your answers are labels from a small set
80%
Valid JSON
Every answer is JSON
95%
Matches the schema
Every answer is a JSON object. The schema is the keys all of them share
90%
Number within 10%
Every answer is a number
70%
Answer in under 2,000 ms
Always offered, but not ticked
None
3
Base model
Under Ready to fine-tune, choose one of the models Valkyrie supports. Each
card shows the model’s size and GPU memory, what it is good for, and the GPU it
trains on. Models marked Tested have completed a fine-tune on Valkyrie.Each card also shows a typical run, such as About 15 to 40 min, $0.05 to
$0.30 for 1,000 rows: the time and cost for 1,000 rows on the Standard
recipe. Your own run’s estimate is in the
Summary. If you opened New fine-tune from the
AI assistant’s card, the model it picked shows
Suggested: and the reason.
Model
Good for
Qwen2.5 7B Instruct
Most tasks: support replies, extraction, a house writing style
Qwen2.5 3B Instruct
Short, well-defined answers at low cost: FAQs, tagging, rewriting
Qwen2.5 14B Instruct
Harder reasoning and longer answers, when 7B is not enough
Llama 3.1 8B Instruct
English-first tasks
Mistral 7B Instruct v0.3
Fast, concise replies; classification and routing
Qwen2.5 0.5B Instruct
Trying the flow end to end; simple labelling
Or switch to Any Hugging Face model to enter another model.
Ready to fine-tune: size, memory, what each model is good for.
4
Recipe
Choose how much training to do:
Recipe
Passes
When to use it
Quick check
1
Find out fast whether your data teaches anything
Standard
Model default
Settings tuned for the base model you picked
Thorough
5
More passes, for larger datasets
Tick Interruptible (cheaper bid box) to train on cheaper GPU time that
can be taken back. Valkyrie bids just above the lowest accepted price, below
the on-demand price. If the machine is taken back, the run resumes from its
last saved checkpoint on a fresh machine, up to 3 times, and the steps since
that checkpoint are trained again. The estimate and the wallet check still
use on-demand prices. See
Interruptible capacity.Open Advanced settings if you need to change the training settings
yourself.
The Summary card on the right shows your examples, base model, recipe, number
of passes, and Capacity (On-demand or Interruptible), with an
Estimate of how long the run takes and what it costs.
The estimate is a range: it counts every row of your file, uses the training speed
measured for the model, adds start-up time, and prices it at current GPU rates. It
updates when you change the examples, model, or recipe. Models outside Ready to
fine-tune have no estimate. Click Start fine-tune to begin.
The estimate: time and cost before you start.
Under the estimate, Balance after shows what your wallet would hold once the
run is paid: what is left after the estimate’s high end, then after its low end
(for example $37.60 to $41.20).Before a run with an estimate starts, Valkyrie checks that the wallet holds the
estimate’s high end plus a 20% buffer. If it doesn’t, Start fine-tune is
disabled and the summary says, for example:
Your balance of $4.10 doesn't cover this run. It needs $6.00: the estimate’s
high end plus a 20% buffer.
Click Add funds next to it to top up, see
Adding funds, or pick a smaller model or
recipe. The same check runs when a fine-tune is started from the
AI assistant, the API, or MCP. A run
without an estimate (a model outside Ready to fine-tune, or a file whose rows
can’t be counted) only needs a positive balance. If a
spend limit would stop the run instead, the
summary says that.
Billed per minute while training. Cancel any time.
After you start, you land on the new version’s page under
Models. It shows the training stages and live progress, and
you can cancel the run from there.
The version page draws a Training loss chart as training logs its progress. See
Training loss.Queued is waiting for a GPU. Initializing covers GPU start-up, model
download, and dataset prep before the first training step. Running is active
training, shown with live progress. Finalizing is the job saving and uploading
its outputs after the last step.Time left starts from the estimate the run was given when it started, as a range
such as “7m to 25m”, and counts down. Once training is underway, it switches to
training’s own estimate, and Finishes around shows the time of day. If the run
passes its estimate, it says Longer than estimated.
Failed: the version explains what went wrong in plain language. Click
Retry to run it again on fresh hardware, or Change settings and try again
to adjust it first.
