[FT] Assistant Response Prefilling

[matmult]

Solution/Feature

Context:
docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/prefill-claudes-response
anthropic.com/news/claude-2-1-prompting

I would love to be able to manually inject assistant responses during eval to perform some ablations. Ideally, I would be able to modify the messages like such:

messages = [
    {
        "role": "user",
        "content": prompt,
    },
        "role": "assistant",
        "content": 'According to ',
    }
]

And it should use this to generate up through assistant's response but not ending it, so the model continues.

Possible alternatives

Adopting a custom chat template could work, but it would be nice to see this feature integrated as I anticipate a lot of researchers will have need for easy access to this.