[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

[ChenhanYu]

Triage Analysis

Field	Value
Classification	Question
Severity	Medium
Complexity	Simple
Auto-fixable	No

Summary

User is asking whether it's recommended to use the same dataset splits for both SFT and QAT fine-tuning, and whether QAT data should match the distribution of RL-tuned models or original SFT data. The issue seeks clarification on best practices for QAT workflows, particularly for models that have undergone RL fine-tuning rather than just SFT.

Root Cause / Approach

The documentation and examples use a single dataset (Daring-Anteater) with different splits for calibration and QAT, which raises questions about whether this is optimal or merely a simplification. Additionally, there is no guidance on handling RL-tuned models, which have different data distributions than SFT models.

Suggested Fix

1. Add explicit documentation in README.md clarifying whether calibration and QAT should use the same or different data distributions. 2) Provide guidance on QAT for RL-fine-tuned models, including whether to use RL-distribution data or original SFT data. 3) Consider adding an additional example or section demonstrating QAT with different dataset distributions. 4) Document empirical findings on accuracy trade-offs when using matching vs. different data splits.

Relevant Files

• examples/llm_qat/README.md
• examples/llm_qat/main.py
• examples/llm_qat/utils.py
• examples/llm_qat/simple_qat_train.py
• examples/llm_ptq/hf_ptq.py
• modelopt/torch/quantization/model_quant.py
• modelopt/torch/quantization/model_calib.py

---

Auto-triaged by pensieve /magic-triage