firefunction-v1

A GPT-4 level function calling model, made by fireworks AI.

Just demonstrate your function as you “declare” it to the model, and it’ll repeat the structure of the call reliably on demand. Extremely high accuracy and reliability for < 5 functions, performs decently for 10 and over as well.

https://huggingface.co/fireworks-ai/firefunction-v1

https://fireworks.ai/blog/firefunction-v1-gpt-4-level-function-calling

Quantized to 4 bits by BNB with:
https://github.com/CharlesMod/quantizeHFmodel

Context length: 2048
(defined by ollama .model file, actual model limit is 32.768kT, limited here for speed.)

EXAMPLE:

ollama run joefamous/firefunction-v1:q3_k
>>> What is the weather in Seattle?
<functioncall>{"name": "get_the_weather", "arguments": {"city": "Seattle", "state": "Washington"}}

>>> TOOL: rainy
<plain>The current weather in Seattle, Washington is rainy.

You should include your list of available functions in the system prompt. The default system prompt contains the following:

You are a helpful assistant with access to functions. Use them if required.

FUNCTIONS: [
  {
    "type": "function",
    "function": {
      "name": "get_the_weather",
      "description": "Get the weather",
      "parameters": {
        "type": "object",
        "properties": {
          "city": {
            "type": "string",
            "description": "The city to get the weather for"
          },
          "state": {
            "type": "string",
            "description": "The state to get the weather for"
          }
        },
        "required": [
          "city",
          "state"
        ]
      }
    }
  }
]