Model card does not include benchmarks, architecture
#1
by nraxl1 - opened
The model card just shows the 30B-A3B variant and just mentions the 2B variant occasionally but some stuff likely does not apply to it. Like I dont think this is post trained on nemotron cascade 30B-A3B
Hi, thank you for raising the question. As stated in the paper:
- The 2B model uses the same SFT recipe as the 30B-A3B model.
- We only apply RL to the 30B-A3B model. The 2B model is the multi-stage SFT checkpoint.
You can find more details in the 2B variant in the paper. We hope the smaller 2B SFT variant could foster easier access to Audex and open research of further post-training including RL.
L0SG changed discussion status to closed