metadata

base_model: Qwen/Qwen2-1.5B-Instruct
datasets:
  - devanshamin/gem-viggo-function-calling
library_name: peft
license: apache-2.0
pipeline_tag: text-generation
tags:
  - trl
  - sft
  - generated_from_trainer
model-index:
  - name: Qwen2-1.5B-Instruct-Function-Calling-v1
    results: []

Qwen2-1.5B-Instruct-Function-Calling-v1

This model is a fine-tuned version of Qwen/Qwen2-1.5B-Instruct on devanshamin/gem-viggo-function-calling dataset.

Updated Chat Template

Note: The template supports multiple tools but the model is fine-tuned on a dataset consisting of examples with a single tool.

The chat template has been added to the tokenizer_config.json.
Supports prompts with and without tools.

chat_template = (
  "{% for message in messages %}"
  "{% if loop.first and messages[0]['role'] != 'system' %}"
  "{% if tools %}"
  "<|im_start|>system\nYou are a helpful assistant with access to the following tools. Use them if required - \n"
  "```json\n{{ tools | tojson }}\n```<|im_end|>\n"
  "{% else %}"
  "<|im_start|>system\nYou are a helpful assistant.\n<|im_end|>\n"
  "{% endif %}"
  "{% endif %}"
  "{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}"
  "{% endfor %}"
  "{% if add_generation_prompt %}"
  "{{ '<|im_start|>assistant\n' }}"
  "{% endif %}"
)

Basic Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Qwen2-1.5B-Instruct-Function-Calling-v1"
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

def inference(prompt: str) -> str:
  model_inputs = tokenizer([prompt], return_tensors="pt").to('cuda')
  generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512)
  generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
  response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
  return response

messages = [{"role": "user", "content": "What is the speed of light?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = inference(prompt)
print(response)

Tool Usage

Basic

import json
from typing import List, Dict

def get_prompt(user_input: str, tools: List[Dict] | None = None):
  prompt = 'Extract the information from the following - \n{}'.format(user_input)
  messages = [{"role": "user", "content": prompt}]
  prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    tools=tools
  )
  return prompt

tool = {
  "type": "function",
  "function": {
    "name": "get_company_info",
    "description": "Correctly extracted company information with all the required parameters with correct types",
    "parameters": {
      "properties": {
        "name": {"title": "Name", "type": "string"},
        "investors": {
          "items": {"type": "string"},
          "title": "Investors",
          "type": "array"
        },
        "valuation": {"title": "Valuation", "type": "string"},
        "source": {"title": "Source", "type": "string"}
      },
      "required": ["investors", "name", "source", "valuation"],
      "type": "object"
    }
  }
}
input_text = "Founded in 2021, Pluto raised $4 million across multiple seed funding rounds, valuing the company at $12 million (pre-money), according to PitchBook. The startup was backed by investors including Switch Ventures, Caffeinated Capital and Maxime Seguineau."
prompt = get_prompt(input_text, tools=[tool])
response = inference(prompt)
print(response)
# ```json
# {
#   "name": "get_company_info",
#   "arguments": {
#     "name": "Pluto",
#     "investors": [
#       "Switch Ventures",
#       "Caffeinated Capital",
#       "Maxime Seguineau"
#     ],
#     "valuation": "$12 million",
#     "source": "PitchBook"
#   }
# }
# ```

Advanced

import re
from enum import Enum

from pydantic import BaseModel, Field # pip install pydantic
from instructor.function_calls import openai_schema # pip install instructor

# Define functions using pydantic classes
class PaperCategory(str, Enum):
  TYPE_1_DIABETES = 'Type 1 Diabetes'
  TYPE_2_DIABETES = 'Type 2 Diabetes'

class Classification(BaseModel):
  label: PaperCategory = Field(..., description='Provide the most likely category')
  reason: str = Field(..., description='Give a detailed explanation with quotes from the abstract explaining why the paper is related to the chosen label.')

function_definition = openai_schema(Classification).openai_schema
tool = dict(type='function', function=function_definition)
input_text = "1,25-dihydroxyvitamin D(3) (1,25(OH)(2)D(3)), the biologically active form of vitamin D, is widely recognized as a modulator of the immune system as well as a regulator of mineral metabolism. The objective of this study was to determine the effects of vitamin D status and treatment with 1,25(OH)(2)D(3) on diabetes onset in non-obese diabetic (NOD) mice, a murine model of human type I diabetes. We have found that vitamin D-deficiency increases the incidence of diabetes in female mice from 46% (n=13) to 88% (n=8) and from 0% (n=10) to 44% (n=9) in male mice as of 200 days of age when compared to vitamin D-sufficient animals. Addition of 50 ng of 1,25(OH)(2)D(3)/day to the diet prevented disease onset as of 200 days and caused a significant rise in serum calcium levels, regardless of gender or vitamin D status. Our results indicate that vitamin D status is a determining factor of disease susceptibility and oral administration of 1,25(OH)(2)D(3) prevents diabetes onset in NOD mice through 200 days of age."
prompt = get_prompt(input_text, tools=[tool])
output = inference(prompt)
print(output)
# ```json
# {
#     "name": "Classification", 
#     "arguments": {
#         "label": "Type 1 Diabetes", 
#         "reason": "The study investigated the effect of vitamin D status and treatment with 1,25(OH)(2)D(3) on diabetes onset in non-obese diabetic (NOD) mice. It also concluded that vitamin D deficiency leads to an increase in diabetes incidence and that the addition of 1,25(OH)(2)D(3) can prevent diabetes onset in NOD mice."
#     }
# }
# ```
# Extract JSON string using regex
output = re.search(r'```json\s*(\{.*?\})\s*```', output).group(1)
output = Classification(**json.loads(_output)['arguments'])
print(output)
# Classification(label=<PaperCategory.TYPE_1_DIABETES: 'Type 1 Diabetes'>, reason='The study investigated the effect of vitamin D status and treatment with 1,25(OH)(2)D(3) on diabetes onset in non-obese diabetic (NOD) mice. It also concluded that vitamin D deficiency leads to an increase in diabetes incidence and that the addition of 1,25(OH)(2)D(3) can prevent diabetes onset in NOD mice.')

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 4
eval_batch_size: 4
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 10
training_steps: 200

Training results

Training Loss	Epoch	Step	Validation Loss
0.4004	0.0101	20	0.4852
0.3624	0.0201	40	0.3221
0.2855	0.0302	60	0.2818
0.2652	0.0402	80	0.2592
0.2214	0.0503	100	0.2463
0.2471	0.0603	120	0.2358
0.2122	0.0704	140	0.2310
0.2048	0.0804	160	0.2275
0.2406	0.0905	180	0.2251
0.2445	0.1006	200	0.2248

Framework versions

peft==0.11.1
transformers==4.42.3
torch==2.3.1+cu121
datasets==2.20.0
tokenizers==0.19.1