Inquire about LLama model training

[liyu0127cs-cyber]

I’d like to ask: when training the Llama-3.2-3B-Instruct model, are there any differences in the configuration of the train.sh? I’ve been using the same script to train the Llama model after changing the prompt template, but the number of model calls drops to zero after only a few dozen steps.

[ViktorAxelsen]

Hello, thanks for your interest.

Aside from regenerating the training data using the Llama-specific prompt template, there should not be any major differences in the train.sh configuration itself. The model’s behavior during training can also be influenced by other factors, such as the composition of the LLM candidate pool and other training settings. Could you please provide a bit more information, such as your training configuration, the candidate pool setup, and any relevant logs?