What does ChatGPT model distillation training data look like?

How is it used in training?

A teacher model produces candidate answers for defined tasks. Rule-based or human review filters them, and accepted answers are paired with their inputs for supervised fine-tuning (SFT) of a student model. Evaluate the student on examples it has never seen. Human-written question–answer pairs can also support SFT, but that does not make them distillation data produced by a particular teacher model.

What does one record contain?

JSONL is common: one record per line. Conversational examples can use a messages array containing roles and content; single-turn tasks often use prompt and completion fields. A separate manifest can record task tags, provenance, review status and split membership. Match the actual field names and types to your training tool.

{"messages":[{"role":"user","content":"Classify this request and return only JSON with an intent field: I received the wrong model and would like an exchange."},{"role":"assistant","content":"{\"intent\":\"exchange\"}"}]}

The example teaches a fixed output field for a customer request. An evaluation can test whether the JSON parses and whether the intent is correct. The same principle scales to harder tasks: give enough context, specify the output and define what passes.

Common tasks and how to check them

Support intent classification

Map de-identified questions to a fixed set of intents. Measure accuracy and recall on underrepresented intents in a held-out set.

Structured extraction

Turn a document or form into defined JSON fields. Test field accuracy, value types and handling of missing information.

Code repair

Pair a minimal reproduction and constraints with a patch. Run independent unit tests instead of grading by resemblance to the teacher’s answer.

How can you tell if the data is useful?

Remove duplicates and near-duplicates, incorrect outputs and sensitive information. Split records by source, task or time where appropriate, so paraphrases of the same question do not leak from training to test. Compare the student before and after fine-tuning on held-out tasks. Measure inference cost or latency separately if those are goals. More records do not guarantee better results.

Four things to verify before buying or accepting a dataset

1. Contents and counts: Inspect actual samples and verify task coverage, languages, unique record counts and the size of each split.
2. Files and schema: Confirm the actual file format, such as JSONL or Parquet, field types, encoding and a reproducible manifest or checksums.
3. Provenance and rights: Identify the sources. For model-generated records, verify the teacher version, applicable provider terms and whether commercial training, redistribution or resale are permitted. Access to an output does not itself grant those rights.
4. Quality and limits: Ask how records were reviewed, what limitations are known and whether evaluation is independent. Put acceptance criteria in the order or delivery documentation.

References

• OpenAI — Model Distillation in the API
• Hugging Face TRL — Dataset formats and types
• Hugging Face Hub — Dataset cards
• OpenAI — Usage Policies