Hi, thanks for the great work on TransArch and OpenDSA!
I'm very interested in the model conversion pipeline and would like to better understand its training dynamics and the quality of the converted model. I was wondering if you could share any experiment logs or results for the conversion process. Specifically:
-
Training curves for the two-stage conversion training — e.g., the loss and grad norm curves during each of the two stages. It would be really helpful to see how stable the training is and roughly how long each stage takes to converge.
-
Benchmark results of the converted model — how does the converted OpenDSA model perform on standard benchmarks compared to the original model?
If these logs or results are already available somewhere, a pointer would be much appreciated. Otherwise, even a brief summary of the training behavior and a few key benchmark numbers would be super helpful. Thanks a lot!
Hi, thanks for the great work on TransArch and OpenDSA!
I'm very interested in the model conversion pipeline and would like to better understand its training dynamics and the quality of the converted model. I was wondering if you could share any experiment logs or results for the conversion process. Specifically:
Training curves for the two-stage conversion training — e.g., the loss and grad norm curves during each of the two stages. It would be really helpful to see how stable the training is and roughly how long each stage takes to converge.
Benchmark results of the converted model — how does the converted OpenDSA model perform on standard benchmarks compared to the original model?
If these logs or results are already available somewhere, a pointer would be much appreciated. Otherwise, even a brief summary of the training behavior and a few key benchmark numbers would be super helpful. Thanks a lot!