Fine-Tune Studio

Upload training data

How this file is labelled

Text field

Label field

Label shape

What the file contains

Can each class be measured?

0.0 ─── shared scale ─── 0.5 hatched = unresolvable (±>0.14)

Findings

Choose what to train

Settings

Run name

Epochs

Seed

Repeat across seeds

Batch size

Learning rate

Class weighting

Calibrate on validation

Precision

Max class weight

Max tokens

Base models

ⓘ Train anyway, despite errors

Project and taxonomy

Project

Name

Shape

Entities / attribute

Sentiment classes

Draft a taxonomy with the model

Taxonomy in force

Editing a description changes the guidelines for the labelling LLM and reviewers; cells checked under the old wording are pooled with the new one until they are re-checked.

Feedback from labellers

Assessment

Project

New assessment dataset

Name

Predictions file

Default cells per stratum

Prompt the LLM should judge with

The model judged EVERY entity on every post - an entity a row does not name is the model's none Judge these cells BLIND

Off is about twice as fast, but measures agreement, which runs higher by the project's acquiescence gap.

Tools

Project Tools that need a project will say so until one is chosen.

Tool

Output

Deploy

1 · Take a run

run head tolerance %

also the speed grid

2 · What was built, and what was refused

Pick a run and build, or open a run that has been built.

3 · Against human, LLM and production

I have validated this teacher — read LLM agreement as accuracy

Pick a run.

Users

Username

Display name

Role

Time and output

Person

Annotator agreement

Project

Local LLM

Name

URL

Labelling model

Translation model

Embedding model

The page failed to load

Sign in

Choose a password

Upload training data ?

How this file is labelled ?

What the file contains ?

Can each class be measured? ?

Languages ?

Findings ?

Choose what to train ?

Settings ?

Runs ?

Results ?

Label review ?

Your work ?

Project and taxonomy ?

Taxonomy in force

Generated prompt ?

The prompt in force — what the next batch will send

Post pool ?

Candidate selection ?

Label a batch ?

Feedback from labellers ?

Verify ?

Coverage ? Exclusions… Print / save as PDF

Synthetic examples ?

Export and train ?

Assessment ?

New assessment dataset ?

Report ? Exclusions… Print / save as PDF

Tools ?

Calibrate against labels you already trust ?

Deploy ?

1 · Take a run ?

2 · What was built, and what was refused ?

3 · Against human, LLM and production ?

Users ?

Time and output ?

Annotator agreement ?

Local LLM ?

Defaults for new runs ?

Activity ?

Upload training data

How this file is labelled

What the file contains

Can each class be measured?

Languages

Findings

Choose what to train

Settings

Runs

Results

Label review

Your work

Project and taxonomy

Generated prompt

Post pool

Candidate selection

Label a batch

Feedback from labellers

Verify

Coverage

Synthetic examples

Export and train

Assessment

New assessment dataset

Report

Tools

Calibrate against labels you already trust

Deploy

1 · Take a run

2 · What was built, and what was refused

3 · Against human, LLM and production

Users

Time and output

Annotator agreement

Local LLM

Defaults for new runs

Activity