Cannot add new roles such as "tool" in ChatML preset

[interstellarninja]

Hermes-2-Pro-Llama-3-8B model uses a new role called "tool" similar to the OpenAI API standard to pass tool results back to the model but the current preset config is too restrictive and doesn't allow adding new roles or non-alternating turns such as multiple system prompts.

Currently when i send function results or errors back to the model with "tool" role i get the BadRequestError:

[
   {
      "role": "user",
      "content": "Your task is to Extract key financial data and ratios from NVDA's SEC filings."
    },
    {
      "role": "assistant",
      "content": "<scratchpad>\nWe need to first find the latest 10-Q and 10-K forms for NVDA and extract relevant information from them.\n</scatchpad>\n\n<tool_call>\n{\"arguments\": {\"data\": \"NVDA|what was last quarter's revenue\"}, \"name\": \"search_10q\"}\n</tool_call>\n"
    },
    {
      "role": "tool",
      "content": "<tool_error>\nXML Parse Error: mismatched tag: line 3, column 2\n</tool_error>\n"
    }
  ]

openai.BadRequestError: Error code: 400 - {'error': "'messages' array must only contain objects with a 'role' field that is either 'user', 'assistant', or 'system'. Got 'tool'."}

There's probably a better way to do this like the Jinja prompt templates but here's what I'm suggesting for the time being allowing multiple input prefixes as a list:

{
  "name": "ChatML",
  "inference_params": {
    "input_prefixes": [
        "<|im_end|>\n<|im_start|>user\n",
        "<|im_end|>\n<|im_start|>tool\n",
        "<|im_end|>\n<|im_start|>agent\n"
    ],
    "input_suffix": "<|im_end|>\n<|im_start|>assistant\n",
    "antiprompt": [
      "<|im_start|>",
      "<|im_end|>"
    ],
    "pre_prompt_prefix": "<|im_start|>system\n",
    "pre_prompt_suffix": "",
    "pre_prompt": "Perform the task to the best of your ability."
  }
}