Rating Answers and the Feedback Loop

StretchGPT gets better because people tell it when it's right and when it's wrong. Every rating you give on an answer can become a training example that improves StretchGPT's proprietary models. This article explains how the feedback loop works and how to give feedback that actually helps.

Why your feedback matters

StretchGPT is designed to learn from real use, not just from generic data. When you rate an answer, StretchGPT captures the exact prompt-and-answer pair and turns it into a review-required training candidate, tagged with the conversation's family. A human then reviews it before it's ever used. This closes the loop: good answers become examples to imitate, and corrected answers become examples of what to do instead.

Thumbs-up: teaching by example

1. When an answer is good, click thumbs-up on that message.
2. StretchGPT records the prompt and the answer as a supervised fine-tuning (SFT) candidate.
3. The candidate enters the training queue marked review required, tagged with the conversation's family.

SFT candidates are examples of the behavior you want more of.

Thumbs-down: teaching by correction

1. When an answer is wrong or weak, click thumbs-down.
2. StretchGPT asks for a correction — the better answer you wish it had given.
3. If you provide one, StretchGPT records a preference (DPO) candidate: the same prompt, with your correction as the chosen answer and the original as the rejected one.
4. If you don't provide a correction, StretchGPT simply notes the rejection and moves on — no training example is created.

DPO candidates teach the model to prefer the better answer over the worse one for that kind of prompt.

A realistic example

In a support conversation, StretchGPT drafts a customer reply that's accurate but too curt. You click thumbs-down, and in the correction box paste a warmer version that keeps the facts. StretchGPT stores a DPO candidate: prompt = the original request, rejected = the curt draft, chosen = your warmer rewrite, tagged to the support family. Later, a reviewer approves it, it's exported into a dataset, and a future support-family model learns to prefer the warmer tone.

What happens after you rate

Your rating doesn't change the model instantly. It becomes a candidate that flows through the governed training pipeline:

1. Review — a human approves or rejects the candidate (see Reviewing and Approving the Training Queue).
2. Dataset export — approved candidates are bundled into a dataset.
3. Training — a job trains a new model or adapter on a GPU node.
4. Evaluation and promotion — the new model is evaluated and only promoted with sign-off (see Managing and Promoting Trained Models).

This is why feedback is safe: nothing you rate is used automatically or without human review.

Tips

• Always add a correction with a thumbs-down. A downvote alone just notes displeasure; a correction creates a real learning example.
• Rate in the right family. Because candidates are tagged by family, feedback improves the specific assistant you were using.
• Correct the substance, not just the style — or both. The clearer your "chosen" answer, the stronger the training signal.
• Thumbs-up genuinely good answers. Positive examples are as valuable as corrections.

FAQ

Does a thumbs-up change the model right away? No. It creates a review-required candidate; a human reviews it before any training happens.

Is my corrected text stored? Yes — as the chosen answer in a preference candidate, tied to the prompt and family, pending review.

Can I rate the same answer twice? Give the rating that reflects your judgment; the most recent, reviewed signal is what matters. Reviewers deduplicate and curate before anything trains.