Reviewing and Approving the Training Queue

StretchGPT improves its own proprietary models through a human-reviewed training pipeline. Nothing trains automatically: candidates are drafted (by teacher generation or from user feedback), a human approves or rejects each one, and only approved candidates are exported into a dataset. This article covers the review workflow. Generating, reviewing, and exporting all require step-up admin authorization.

Where candidates come from

Training candidates enter the queue two ways:

1. User feedback. A thumbs-up becomes a supervised (SFT) candidate; a thumbs-down with a correction becomes a preference (DPO) candidate. Both arrive marked review required (see Rating Answers and the Feedback Loop).
2. Teacher generation. Operators ask a teacher model to draft candidates on demand — SFT examples for a family and topic, or a DPO example for a given prompt and rejected answer.

Every candidate is tagged with its family and task type (SFT or DPO) and starts in a review required state.

Generating candidates with the teacher

Supervised (SFT)

1. In the training console, choose Generate SFT candidates.
2. Pick a family (defaults to Core), a topic, and how many to draft (count).
3. Complete the step-up authorization. The teacher drafts the candidates into the queue as review required.

Preference (DPO)

1. Choose Generate DPO candidate.
2. Provide the family, the prompt, and the rejected answer you want to improve on.
3. Authorize. The teacher drafts a preference candidate for review.

Reviewing candidates

1. Open the Training candidates list. Filter by family, task type, or review status to focus your review.
2. Open a candidate to see the prompt, the response (or chosen/rejected pair for DPO), how it was generated, and its provenance.
3. Decide:
- Approve if it's a good example.
- Reject if it's wrong, low-quality, or off-policy.
4. Add notes explaining your decision — useful for auditing and for other reviewers.
5. Submit. Your identity and the decision are recorded on the candidate.

Use the candidate stats view to see the shape of your queue (counts by family and status) so you know what's waiting.

Exporting an approved dataset

Once you have a batch of approved candidates:

1. Choose Export dataset.
2. Set the family, the kind (SFT or DPO), and an optional title.
3. Authorize. StretchGPT bundles the approved candidates for that family and kind into a dataset, recording who created it.

Only approved candidates are exported — rejected and still-pending ones are left out. The dataset is what a GPU node later trains on (see Managing and Promoting Trained Models).

A realistic example

Your support team has been rating answers for a week, so the queue has 40 support-family candidates. You filter to family: support, status: review required and work through them: approving the clean examples, rejecting three that leaked internal jargon, and leaving notes on the borderline ones. You also generate 10 fresh SFT candidates on a topic you know is thin. When you've approved ~35 good examples, you export a support / SFT dataset titled "Support tone v2." It's now ready to train a better support-family model.

Tips

• Review by family. Filtering keeps you in one context and makes decisions faster and more consistent.
• Reject freely. A smaller, cleaner dataset trains a better model than a large, noisy one.
• Always leave notes. They're your audit trail and help other reviewers stay consistent.
• Balance sources. Mix real user-feedback candidates with teacher-generated ones for coverage.

Troubleshooting

My export is empty or missing candidates. Export only includes approved candidates for the chosen family and kind. Approve more, or check that you selected the right family/kind.

A candidate I approved isn't in the dataset. Confirm its family and task type match the export you ran — an SFT export won't pull DPO candidates, and vice versa.

Generation or review was rejected as unauthorized. These are gated writes needing the training-manage scope. Complete the step-up authorization and retry.