How should this method be optimized to target multiple test datasets, including GSM8K and MMLU?

[anaivebird]

Is this method only suitable for fine‑tuning on GSM8K and not for optimizing across all metrics? We’ve found that while GSM8K performance remains strong, other metrics—like MMLU—drop off sharply.

[YouCanLX]

Could you please share the KVTuner configuration and models tested on MMLU like datasets?

Our method is a general mixed precision KVCache quantization mode and precision tuning framework. Therefore, it can be applied to optimize other metrics.