What format are SFT exports?

Standard JSON training pairs: input (the original AI message) and target (the human-approved correction). Compatible with common fine-tuning frameworks. No proprietary format, no lock-in.

Do SFT exports include all messages?

Only approved corrections from needs_fix and blocked items. safe_to_deploy messages aren't included because they don't represent corrections — there's nothing for the model to learn from a message that was already right. The training data is exclusively from human-reviewed, human-corrected, human-approved items.

How much training data do I need?

Less than you'd think. Even 50-100 high-quality correction pairs can measurably improve model performance for specific tasks like tone, compliance language, or product accuracy. Most teams accumulate that volume within the first month of running messages through the evaluation process.

SFT Export for AI Training Data | Bookbag Intelligence

What It Means

Key Insight

Every human correction in the AI QA & Evaluation Platform is already a training example waiting to happen. SFT export just packages it for your model to learn from.

SFT (Supervised Fine-Tuning) export takes the corrections your human reviewers make in the AI QA & Evaluation Platform and packages them as training data. The format is simple: each pair has the original AI-generated message (the input your model produced) and the human-approved correction (what it should have produced). Feed these pairs to your fine-tuning pipeline, and you're directly teaching the model what 'good' looks like — not in the abstract, but for your specific brand, your compliance requirements, your tone, your product facts. It's the most straightforward form of training data because there's no ambiguity: the human said 'this is wrong, here's what's right.' The model learns the difference.

Why It Matters

Here's the training data flywheel: your AI generates messages. The AI QA & Evaluation Platform catches the ones that need fixing. Humans correct them. Those corrections become SFT pairs. You retrain the model. The AI produces better messages. Fewer corrections needed. Higher safe_to_deploy rates. Lower review costs. The flywheel is powered by the corrections your team is already making — SFT export just captures the value.

How Bookbag Helps

Approved-only export

Only corrections that were reviewed and approved become training data. No unapproved or draft corrections contaminate your training set.

Full provenance metadata

Every training pair includes which reviewer made the correction, which rubric applied, and when — so you can trace any training example back to its source.

Standard format output

JSON pairs with input/target structure, compatible with common fine-tuning frameworks. Plug directly into your training pipeline.

SFT Export

What It Means

Why It Matters

How Bookbag Helps

Approved-only export

Full provenance metadata

Standard format output

Related Terms

Frequently Asked Questions

Related Resources

Solutions

Compare

See how Bookbag works