MOSS-Image-SFT training pipeline transparency and methodology.
Reproducibility, methodological clarity, and advanced fine-tuning techniques (SFT vs. RL/RLHF).
This issue exposes a critical transparency deficit in the MOSS-Image-SFT training pipeline documentation. Developers require granular details on dataset composition, loss functions, learning rates, and training steps to understand model behavior, reproduce results, and build upon the foundation. ...
MOSS-Image-SFT
fine-tuned
MOSS-Image-Base
curated data
SFT dataset composition
View Technical Brief