Model TrainingComing soon to Scalata Enterprise
Teach a model to answer the way your team does
Turn your best conversations, documents and feedback into a model of your own. Prove it is better before anyone relies on it, and keep every step governed, costed and on the record.
Your best answers are already written. They are in the conversations your team has every day — precise, in your tone, about your products. Model Training turns them into a model that answers like that from the first message.
How it works
From conversation to model in six steps
No machine-learning team needed. Every step shows what it will do before it does it, and what it did afterwards.
- Step 01
Pick the conversations worth learning from
Choose the chats where your team got it right. Each reply you keep becomes something the model learns to write; the rest stays as context.
- Step 02
Build a dataset from every source
Add generated conversations for the cases you rarely see, import public datasets, attach documents, and fold in what people liked and disliked in chat.
- Step 03
Prepare it, safely
Errors, duplicates and canned replies are removed, personal details are redacted, long chats are trimmed to fit, and a test set is put aside that training never sees.
- Step 04
Train, with the cost in front of you
Pick a method and a base model, keep the suggested settings or tune them, and see the estimated cost before you start. Progress, curves and an email when it is done.
- Step 05
Prove it beat the original
Your new model and the one it started from answer the same held-out questions. A judge compares them blind, and you read why each answer won.
- Step 06
Try it, then put it to work
Chat with the trained model for as long as you choose, with the hourly cost shown up front and nothing left running afterwards. When you are happy, it is ready for your team.
Ways to learn
Every way a model can learn, in one place
The methods differ in what the model learns from. Start with the one that matches the data you have — and build on a model you already trained rather than starting over.
Supervised
- Learns from
- Your best replies. It is shown the question and the answer you want, and learns to write that answer.
- Best for
- Tone, format and the way your team phrases things. The first method to try.
Your data
Data from wherever it already lives
Real conversations
The chats your team already has in Scalata — the most faithful picture of how you answer.
Generated conversations
Describe the assistant, review a plan of scenarios, and get varied chats — including the hard and out-of-scope ones.
Public datasets
Bring in an open dataset; its format is recognised and the right columns are suggested for you.
Documents
Manuals, help pages and reports, uploaded or picked from the files and sources you have connected.
Chat feedback
Thumbs up become replies to learn from; a liked and a disliked answer become a preference pair.
Evaluation
Proof, not promises
A falling training curve says the model fits your examples. It does not say the answers got better. Every run is compared with the model it started from, on questions it never saw — and you can rank several models side by side.
Share of held-out questions the new model answered better.
Measured from blind human review, so you know how far to trust the scores.
Governed by default
Built for organisations that have to answer for their models
Personal details stay out
Emails, phone numbers, card numbers and IDs are found and replaced before anything is trained, so the model learns the pattern, not the person.
A test the model never saw
A share of conversations is set aside before training starts and used only to judge the result, so a good score means a better model rather than a good memory.
Costs before, during and after
An estimate before you start, the running figure while it trains, and the final amount when it ends — for training, evaluation and every test chat.
Checked by people, not just by a model
Send any evaluation to a review queue. Reviewers judge blind, their agreement with the judge is measured, and their corrections become training data.
Nothing is lost
Runs you no longer need are archived rather than deleted, so you can always see what was tried, how it scored and which model it came from.
Permissioned like everything else
Training will sit inside your Scalata Enterprise deployment, under the same roles, grants and record as the rest of the platform.
Learn the concepts
New to model training? Start here
A model of your own, inside the platform you already run
Model Training is coming soon to Scalata Enterprise licences. Talk to us now about early access — and about the conversations you would teach it first.