[Question] Is it recommended to use the same examples for SFT and QAT?

[DhruvBhatia0]

Quantizing a model to NVFP4. The model was SFT'd then RL fine-tuned. Planning to do QAT to recover accuracy after PTQ.

Looking at examples/llm_qat/, the same Daring-Anteater dataset is used across all three steps (SFT → PTQ
calibration → QAT), just different splits. A few questions:

1. Calibration vs QAT data — Is using the exact same samples for both SFT and QAT fine-tuning
   intentional/recommended, or just a simplification for the example?

2. RL-tuned models — My model's final training stage was RL, not SFT. Should QAT fine-tuning data
   match the RL distribution (tool calls, reasoning traces, environment feedback), or the original SFT data, or
   does it not matter much?

Who can help?

@kevalmorabia97 @sugunav14