How to embed system prompts into the llms?

[JamesKnight0001]

How to embed system prompts into the LLM? Like without manually doing it like this:

history = []
history.append({"role": "system", "content": "You are a...."})

Because the LLM I am working on is sometimes forgetting the system prompt and I don't want to re-add the system prompt. And If I train it with the instructed dataset I don't want it to start saying its instructions.

[cg123]

You could alter the prompt template in tokenizer_config.json to always include a system message. There aren't any options for doing this in mergekit, though - it just gets copied right through.

[JamesKnight0001]

You could alter the prompt template in tokenizer_config.json to always include a system message. There aren't any options for doing this in mergekit, though - it just gets copied right through.

Thanks