Generic benchmarks
A public benchmark can't tell you whether a reply fits your character, tone, or rules.
How we train
Train a model on the replies your team approves.
We prepare conversation data, train an adapter for your chosen model, and test it against your best prompt. Your team gets the dataset, results, and settings to run it. We work with character apps, AI companions, and voice agents.
Start with the replies you want to change.
If better prompts haven't fixed the same recurring mistake, fine-tuning may help. First, your reviewers need to agree what a good reply looks like.
Annotation without context
Reviewers need the preceding turns and relevant tool state to approve a useful reply.
Weak comparisons
The trained model needs to beat the prompt you would actually use, not just an empty one.
Lost human edits
Reviewer rewrites can become training candidates, but they need approval and a held-out split.
The dataset, adapter, and evidence behind it
We choose examples with your reviewers and test whether the trained model is worth using. If a prompt or tool fix works just as well, we'll recommend it.
Conversation records
We assemble the messages, tool results, and version details from your existing exports.
Human-reviewed next turns
Approved replies tied to a rubric, with separate training and held-out conversation sets.
Adapter and serving checks
Weights for the agreed base model, tested with the prompt, quantization, and runtime you plan to use.
Design the training set before you train
Choose one task, decide which conversations to include, and write clear review criteria. Keep a source reference for every approved reply.
Choose the examples
Focus on a recurring problem. Include difficult cases and good replies that show the tone you want.
Agree review rules
Decide what reviewers can change, who resolves disagreements, and what evidence they need to approve a reply.
Keep the source
Store the original conversation, correction, reviewer, and policy version together so you can check or remove an example later.
Build approved examples from real conversation logsTest whether the model change helps
What you get
We agree file formats and usage rights before work starts. Data preparation and testing are part of the project.
Training records
Approved JSON or CSV examples, review rules, and a list of the cases covered for one task.
Separate test set
Save conversations for testing and keep them out of the training data. Record which examples are in each set.
Source and review history
Each record includes the original conversation, approved reply, review notes, version, and rules for reuse.
Consistent reviews
Have reviewers score the same examples first. Resolve disagreements with the person responsible for approving the model's replies.
Adapter and test results
If training goes ahead, receive the agreed weights, comparison results, capability tests, and settings to run the model.
Data handling
Remove unnecessary sensitive information. Agree who can access the data, how it may be used, and when it will be deleted before transfer.
Free template
Get the sample dataset
See how to record the original conversation, the mistake, the approved reply, and the reviewer's notes.
This template uses made-up examples, not client data. Adapt it to your policies, privacy requirements, and approval process.
Who this is for
- Founders, CTOs, and ML leads building character, companion, or voice products
- Teams whose agent still gives unwanted replies after prompt improvements
- Products with conversation data they have permission to use for training
- Teams with reviewers who can define and approve the behavior they want
Custom Model Sprint deliverables
Review
Review the conversations, group recurring mistakes, and agree how to score replies.
Train
Prepare approved replies and train an adapter if the review shows it's worth testing.
Evaluate
Compare with the best prompt on separate test conversations, check for new failures, and recommend the next step.
Check for new failures before deployment
Agree the tests before training. Keep test conversations out of the training set and run the model in the software you plan to use. Check the tasks it already did well. Keep changing facts and policies in documents or tools it can look up.
Limitations
- Outdated examples or examples from the wrong task can make replies worse. Check retrieval filters and remove old records.
- Review approved datasets when policies, tools, or products change. Remove examples that no longer apply.
- Logs can contain personal information. Remove unnecessary details before review.
- If reviewers can't agree on an answer, keep it out of the approved dataset until the disagreement is resolved.
How we test changes
We review real conversations with your team and agree what the agent should do differently. Approved corrections can become training data, test cases, or examples the agent looks up. We compare each change with your best existing setup on separate test cases, then check for new failures in the software you plan to deploy.
References:
Common questions
What does the work include?
We choose examples with your reviewers and test whether the trained model is worth using. If a prompt or tool fix works just as well, we'll recommend it.
Do you fine-tune our model?
If it's worth testing after we review the problem. We try your best prompt and consider better examples or tool fixes first. We test an adapter on cases it wasn't trained on and check for new failures. The contract sets delivery and usage rights; the base model's license still applies.
What do we need to start?
Logs or conversation exports for one task, examples of good results, and a reviewer who can approve corrections. We agree data handling before you send anything. Remove unnecessary personal details.
How is success measured?
We test the existing and changed agent on the same task. Where the data is available, we compare completed tasks, human edits, handoffs, response time, and cost.
Related
Get the template or discuss your training data.
Use the sample to plan your dataset. If you want help, book a call about one task your agent needs to handle better.