Open MDK-L opened 6 months ago
I found these two models on OpenAI's LLM, and they exhibit significant differences in performance. I'm wondering why? Could it be due to differences in floating-point precision?
Hello @MDK-L! Thank you for your interest in our AlpaGasus2-QLoRA.
As you asked, you can see that two AlpaGasus2-QLoRA models have been uploaded to the Open LLM Leaderboard. Among the two models, AlpaGasus-2-13b-QLoRA-pipeline is the model created for testing in the initial experiments. The final model created after additional modifications was AlpaGasus-2-13b-QLoRA-merged. The difference in performance between the two models is probably due to differences in the learning process.
I hope this answer helps resolve your questions, and if you have any additional questions, please feel free to ask.
Thank you for your answer, but I still have some doubts. When I choose the floating point precision in Open LLM Leaderboard, the 4-bit corresponds to a "pipeline" model with lower accuracy, while float16 and bfloat16 correspond to two "merged" models with similar higher accuracy. This confuses me. Shouldn't the Qlora model be 4-bit?