tomviner/llm-claude

Human: / Assistant: formatting

Opened this issue · 0 comments

tomviner commented

llm-claude currently follows the advice at https://docs.anthropic.com/claude/reference/getting-started-with-the-api#prompt-formatting to:

when using the API you must format the prompts like:
\n\nHuman: Why is the sky blue?\n\nAssistant:

For reference, these are the prompt constants from https://github.com/tomviner/anthropic-sdk-python/blob/d9b7e30b26235860fe9e7e3053171615173e32ca/src/anthropic/_constants.py#L1

HUMAN_PROMPT = "\n\nHuman:"

AI_PROMPT = "\n\nAssistant:"

However, it's allowed to include text starting the AI assistant's answer for it, which isn't currently possible with llm-claude.

On the Ask Claude for rewrites docs page, they explain how starting the AI answer with special text, can assist with formatting it's response:



Human: {prompt listing some text and requesting a rewrite according to some rules}
Please put your rewrite in <rewrite></rewrite> tags.

Assistant: <rewrite>

this starts the assistant already in rewrite mode.

Couple example of starting the AI response here, when you only want a specific format:

And another starting in <thinking> mode: https://docs.anthropic.com/claude/docs/roleplay-dialogue#complex-customer-support-agent

Another:

Assistant: Here is a summary of the document:

from https://docs.anthropic.com/claude/docs/claude-is-hallucinating#document-summary

Why?

Claude has been trained and fine-tuned using RLHF (reinforcement learning with human feedback) methods on \n\nHuman: and \n\nAssistant: data like this, so you will need to use these prompts in the API in order to stay “on-distribution” and get the expected results. It's important to remember to have the two newlines before both Human and Assistant, as that's what it was trained on.

However, I don't think llm-claude should leave the human/assistant prefixes complete up to the end user, because their presence is actually enforced at the API level:

$ curl -s --request POST \
     --url https://api.anthropic.com/v1/complete \
     --header "anthropic-version: 2023-06-01" \
     --header "content-type: application/json" \
     --header "x-api-key: $ANTHROPIC_API_KEY" \
     --data '
{
  "model": "claude-instant-1",
  "prompt": "\n\nHuman: 1+1",
  "max_tokens_to_sample": 10
}
'
{"error":{"type":"invalid_request_error","message":"prompt must end with \"\n\nAssistant:\" turn"}}

From playing around like this, the rules appear to be:

  • prompt must start with {HUMAN_PROMPT}, well almost. Only 1 new line needed at the start. So "\nHuman:"
  • prompt may contain multiple human and AI prompt prefixes. This is how continued conversation are sent.
  • prompt must include at least one {AI_PROMPT} that doesn't have a {HUMAN_PROMPT} after it.

These rules are similarly listed at https://docs.anthropic.com/claude/docs/prompt-troubleshooting-checklist#the-prompt-is-formatted-correctly