File size: 3,579 Bytes
980c77d 7498e73 980c77d c0e8201 7498e73 c0e8201 7498e73 980c77d 7498e73 980c77d 7498e73 980c77d 7498e73 980c77d 7498e73 980c77d 7498e73 980c77d 7498e73 980c77d 7498e73 980c77d 7498e73 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 |
---
language:
- en
license: apache-2.0
library_name: transformers
tags:
- function-calling
- xkcd
- gemma
- unsloth
- peft
- lora
base_model: google/functiongemma-270m-it
datasets:
- olivierdehaene/xkcd
pipeline_tag: text-generation
---
# XKCD FunctionGemma
A fine-tuned version of [google/functiongemma-270m-it](https://huggingface.co/google/functiongemma-270m-it) for XKCD comic search function calling.
## Model Description
This model was fine-tuned to generate structured function calls for searching XKCD comics. Given a natural language query about comics, it outputs a properly formatted tool call that can be parsed and executed.
**Base model:** `google/functiongemma-270m-it`
**Fine-tuning method:** LoRA via Unsloth (1.4% trainable parameters)
**Training data:** 2,630 examples from [olivierdehaene/xkcd](https://huggingface.co/datasets/olivierdehaene/xkcd)
**Training time:** ~8 minutes on T4 GPU
## Usage
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import json
import re
# Load model
model = AutoModelForCausalLM.from_pretrained(
"gnumanth/xkcd-functiongemma",
device_map="auto",
torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("gnumanth/xkcd-functiongemma")
# Define tools
TOOLS = [{
"type": "function",
"function": {
"name": "search_xkcd",
"description": "Search XKCD comics by topic",
"parameters": {
"type": "object",
"properties": {"query": {"type": "string"}},
"required": ["query"]
}
}
}]
# Generate function call
messages = [{"role": "user", "content": "Find xkcd about programming"}]
text = tokenizer.apply_chat_template(messages, tools=TOOLS, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128, do_sample=False)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
# Output: <start_function_call>call:search_xkcd{"query": "programming"}<end_function_call>
```
## Parsing Function Calls
```python
def parse_function_call(output: str) -> dict | None:
"""Extract function name and arguments from model output."""
match = re.search(r'call:(\w+)\s*\{(.+)\}', output, re.DOTALL)
if not match:
return None
func_name = match.group(1)
args_raw = match.group(2).strip()
# Handle double braces from training format
args_raw = re.sub(r'^\s*\{', '', args_raw)
if args_raw.endswith('}'):
args_raw = args_raw[:-1]
try:
return {"function": func_name, "arguments": json.loads('{' + args_raw + '}')}
except json.JSONDecodeError:
return None
# Usage
call = parse_function_call(response)
# {'function': 'search_xkcd', 'arguments': {'query': 'programming'}}
```
## Training Details
- **Epochs:** 1
- **Batch size:** 2 (with 4 gradient accumulation steps)
- **Learning rate:** 2e-4
- **LoRA rank:** 16
- **Target modules:** q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- **Final loss:** 0.281
## Limitations
- Only trained for XKCD search queries
- May produce double braces in output (handled by parser above)
- Small model (270M params) - limited reasoning capability
## License
Apache 2.0 (same as base model)
## Links
- [Training notebook](https://github.com/hemanth/notebooks/blob/main/notebooks/functiongemma_xkcd_finetune.ipynb)
- [Base model](https://huggingface.co/google/functiongemma-270m-it)
- [XKCD dataset](https://huggingface.co/datasets/olivierdehaene/xkcd)
|