What if use LLM without default system prompt？
What if we use an LLM without the default system prompt? Will there be any degradation?@alexrs
The default system prompt eats up extra tokens (even though prefilling is cheap and fast). Is there any point in training with this prompt? I feel like it stops the model from being used in more diverse ways. @alexrs . Is this prompt added during the RF stage or the SFT stage?
@MRU4913 We support different system prompts to achieve different behavior (i.e interactive and non-interactive, different levels of safety, RAG and tools, etc) -- https://docs.cohere.com/docs/command-a-hf
It does use extra tokens in the input, but the model is trained to support a context length of 256K, so there should still be enough room there for most use cases!
feel like it stops the model from being used in more diverse ways
You can definitely experiment with no system prompt and custom system prompts if your use case can benefit from it (and you're encouraged to share results if you do so!)