Fine-tune a model on
your own conversations
Turn the threads your team answered well into a training set, tune an open-weight model on them, and point an agent at the result — so the tone you keep re-explaining in a prompt lives in the weights instead.
Dataset, run, agent
Three steps, all inside the workspace the model will work in.
Build a dataset
Choose the conversations worth learning from and turn them into a training set. Review what went in before anything is sent for training.
Pick a base and train
Choose an open-weight base model — the catalogue shows size and training rate for each — set the hyperparameters or take the defaults, and start the run.
Point an agent at it
The finished adapter shows up as a selectable model. Run it beside the model you use today, on the same conversations, and keep the better one.
Build a dataset
Choose the conversations worth learning from and turn them into a training set. Review what went in before anything is sent for training.
Pick a base and train
Choose an open-weight base model — the catalogue shows size and training rate for each — set the hyperparameters or take the defaults, and start the run.
Point an agent at it
The finished adapter shows up as a selectable model. Run it beside the model you use today, on the same conversations, and keep the better one.
Fine-tuning without a separate stack
No provider console, no reserved capacity, no contract — the same credit balance as everything else.
Train on work you already did
Your best conversations are the training set. Pick the threads your team answered well and build a dataset out of them, instead of inventing examples for a model to imitate.
A smaller model that behaves
Once the tone and the vocabulary are in the weights, they stop costing you prompt tokens on every call — which usually means a smaller, faster model doing work that needed a large one.
Priced before you start
Every base model shows its parameter count and its training rate per million tokens up front. Training is debited from the same credit balance as everything else — no separate contract, no reserved capacity.
A dropdown, not a migration
A finished adapter appears as a model your agents can select, next to the hosted providers. Point one agent at it, compare, and keep whichever answers better.
Jobs you can watch
Datasets, training runs and finished adapters each have their own list, with the state, the base model and the cost of every run visible rather than buried in a provider console.
Gated on its own
Training data is made of real customer conversations, so the whole surface carries its own permission — including the candidate list — separate from the rest of the workspace.
What the pipeline supports
From the conversations that go in to the adapter your agents call.
Datasets
- Built from your own conversations
- Candidate threads listed for review before inclusion
- A dataset detail view of what will be trained on
- Reusable across several training runs
Base models
- Open-weight families: Llama, Qwen, Mistral, Gemma
- DeepSeek, GLM, Nemotron, Kimi and gpt-oss
- Parameter count shown per model
- Training rate per million tokens shown before you start
- Catalogue read live from the training backend
Training runs
- LoRA adapters, with sensible defaults for every hyperparameter
- Job state and history per run
- Cost recorded against the run
- Failures surfaced, not silently retried forever
Serving
- Adapters appear as a model provider in the agent picker
- Billed at the base model inference rate
- Usable by any agent in the workspace
- Switchable per agent, per node
Control
- Its own permission key, with training a separate action
- Candidate conversations gated behind the same key
- Datasets never leave your workspace
- Not pooled with, or used to train, anything else
Cost
- Same credit balance as the rest of the platform
- No subscription, no per-seat charge
- Rate visible before the run, cost recorded after
What a tuned model changes
Where the cost and the inconsistency actually go.
A 2,000-token system prompt re-explaining your tone on every single call
Tone and vocabulary in the weights, paid for once
Reaching for the biggest model because the small one will not stay on brand
A tuned small model that behaves, at a fraction of the per-call cost
Writing synthetic training examples for a domain you already have transcripts of
A dataset built from the conversations your team already answered well
Training in a provider console, disconnected from where the model is used
Dataset, run and served adapter in the same workspace as the agents
Questions about fine-tuning
What teams check before they train on customer conversations.
A prompt tells a model what to do; a fine-tune teaches it how you do it. Once the tone, the product vocabulary and the shape of a good reply are in the weights, you stop paying for them in every prompt — which usually means a smaller, cheaper, faster model doing work that previously needed a large one.
From conversations you have already had. You pick the threads worth learning from — the ones your team answered well — and build a dataset out of them, rather than writing training examples from nothing. Because those conversations contain real customer messages, the whole surface is permission-gated separately from the rest of the workspace.
Open-weight models only, since a hosted model like Claude or GPT cannot be LoRA-tuned. The catalogue is read live from the training backend and covers the usual open families — Llama, Qwen, Mistral, Gemma, DeepSeek, GLM and others — each shown with its parameter count and its training rate before you start.
Training is priced per million tokens at the rate shown for the base model you picked, and it is debited from the same credit balance as everything else. Once trained, an adapter is billed at its base model’s inference rate. There is no subscription and no per-seat charge — see the pricing page for how usage billing works.
It appears as a model your agents can select, alongside the hosted providers. Point one agent at it, compare it against what you were using, and keep whichever answers better — the switch is a dropdown, not a migration.
No. A dataset built in your workspace trains an adapter belonging to your workspace. It is not pooled, not shared with other tenants, and not used to improve any model outside it.
Train a model that already sounds like you
Build a dataset from conversations you have already had. Free to start — you only pay for the training run itself.