metadata

library_name: transformers
datasets:
  - hypervariance/function-calling-sharegpt

Model Card for Model ID

Gemma 2B function calling. google/gemma-2b-it finetuned on hypervariance/function-calling-sharegpt.

Usage

from transformers import AutoModelForCausalLM , AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("rodrigo-pedro/gemma-2b-function-calling", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("rodrigo-pedro/gemma-2b-function-calling", trust_remote_code=True, device_map="auto")

inputs = tokenizer(prompt,return_tensors="pt").to(model.device)

outputs = model.generate(**inputs,do_sample=True,temperature=0.1,top_p=0.95,max_new_tokens=100)

print(tokenizer.decode(outputs[0]))

You can also use sharegpt formatted prompts:

from transformers import AutoModelForCausalLM , AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("rodrigo-pedro/gemma-2b-function-calling", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("rodrigo-pedro/gemma-2b-function-calling", trust_remote_code=True, device_map="auto")

chat = [
  {
      "from": "system",
      "value": "SYSTEM PROMPT",
  },
  {
      "from": "human",
      "value": "USER QUESTION"
  },
]

prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs,do_sample=True,temperature=0.1,top_p=0.95,max_new_tokens=100)

print(tokenizer.decode(outputs[0]))

Prompt template

You are a helpful assistant with access to the following functions. Use them if required -
{
    "name": "function name",
    "description": "function description",
    "parameters": {
        "type": "type (object/number/string)",
        "properties": {
            "property_1": {
                "type": "type",
                "description": "property description"
            }
        },
        "required": [
            "property_1"
        ]
    }
}

To use these functions respond with:
<functioncall> {"name": "function_name", "arguments": {"arg_1": "value_1", "arg_1": "value_1", ...}} </functioncall>

Edge cases you must handle:
 - If there are no functions that match the user request, you will respond politely that you cannot help.

User Question:
USER_QUESTION

Function calls are enclosed in <functioncall> </functioncall>.

The model was trained using the same delimiters as google/gemma-2b-it:

<bos><start_of_turn>user
Write a hello world program<end_of_turn>
<start_of_turn>model

Use <end_of_turn> stop sequence to prevent the model from generating further text.