File size: 8,581 Bytes
6e013ad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
---
language: ko
license: mit
tags:
  - function-calling
  - korean
  - banking
  - on-device
  - onnx
  - int8
  - webgpu
base_model: google/functiongemma-270m-it
---

# TransferFunctionGemma

ํ•œ๊ตญ์–ด ์ž์—ฐ์–ด ์ด์ฒด ๋ช…๋ น์„ ๊ตฌ์กฐํ™”๋œ function call๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๊ฒฝ๋Ÿ‰ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.

## Model Description

TransferFunctionGemma๋Š” [google/functiongemma-270m-it](https://huggingface.co/google/functiongemma-270m-it)๋ฅผ ํ•œ๊ตญ์–ด ๊ธˆ์œต ์ด์ฒด ๋„๋ฉ”์ธ์— ๋งž๊ฒŒ full fine-tuningํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ž์—ฐ์–ด ์ด์ฒด ๋ช…๋ น์„ ๋ถ„์„ํ•˜์—ฌ 4์ข…๋ฅ˜์˜ function call JSON์œผ๋กœ ๋ณ€ํ™˜ํ•ฉ๋‹ˆ๋‹ค.

ONNX INT8 ์–‘์žํ™”๋ฅผ ํ†ตํ•ด ์•ฝ 418MB๋กœ ๊ฒฝ๋Ÿ‰ํ™”๋˜์—ˆ์œผ๋ฉฐ, Transformers.js + WebGPU๋ฅผ ํ†ตํ•ด ๋ธŒ๋ผ์šฐ์ €์—์„œ ์ง์ ‘ ์ถ”๋ก ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์„œ๋ฒ„ ํ†ต์‹  ์—†์ด 100% ํด๋ผ์ด์–ธํŠธ ์‚ฌ์ด๋“œ์—์„œ ๋™์ž‘ํ•ฉ๋‹ˆ๋‹ค.

### Supported Functions

| Function | ์„ค๋ช… | ํ•„์ˆ˜ ์ธ์ž |
|----------|------|-----------|
| `execute_transfer` | ์ˆ˜์ทจ์ธ์—๊ฒŒ ๊ธˆ์•ก์„ ์ด์ฒดํ•ฉ๋‹ˆ๋‹ค | `recipient`, `amount` |
| `query_history` | ์ด์ฒด ๋‚ด์—ญ์„ ์กฐํšŒํ•ฉ๋‹ˆ๋‹ค | (์—†์Œ) |
| `summarize_history` | ์ด์ฒด ๋‚ด์—ญ์„ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค | `period` |
| `confirm_transfer` | ๋Œ€๊ธฐ ์ค‘์ธ ์ด์ฒด๋ฅผ ํ™•์ธ/์ทจ์†Œ/์ˆ˜์ •ํ•ฉ๋‹ˆ๋‹ค | `action` |

---

## Intended Use

### Primary Use Cases

- **๋ธŒ๋ผ์šฐ์ € ๊ธฐ๋ฐ˜ ์ด์ฒด ๋ฐ๋ชจ**: URL ์ ‘์†๋งŒ์œผ๋กœ ์ž์—ฐ์–ด ์ด์ฒด ๊ธฐ๋Šฅ์„ ์ฒดํ—˜
- **ํฌํŠธํด๋ฆฌ์˜ค ๋ฐ๋ชจ**: ๊ธฐ์ˆ  ๋ฉด์ ‘๊ด€/์ฑ„์šฉ ๋‹ด๋‹น์ž์—๊ฒŒ ์˜จ๋””๋ฐ”์ด์Šค AI ์—ญ๋Ÿ‰ ์‹œ์—ฐ
- **์˜จ๋””๋ฐ”์ด์Šค AI ๋ ˆํผ๋Ÿฐ์Šค**: FunctionGemma fine-tuning + ๋ธŒ๋ผ์šฐ์ € ๋ฐฐํฌ ํŒŒ์ดํ”„๋ผ์ธ ์ฐธ๊ณ 

### Out-of-Scope Use

- ์‹ค์ œ ๊ธˆ์œต ๊ฑฐ๋ž˜์— ์‚ฌ์šฉํ•˜๋ฉด ์•ˆ ๋ฉ๋‹ˆ๋‹ค (์ด ๋ชจ๋ธ์€ ๋ฐ๋ชจ ์ „์šฉ์ž…๋‹ˆ๋‹ค)
- ์ด์ฒด ์™ธ์˜ ๊ธˆ์œต ์—…๋ฌด(๋Œ€์ถœ, ํˆฌ์ž, ๋ณดํ—˜ ๋“ฑ)์—๋Š” ํ•™์Šต๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค
- ์˜์–ด ๋“ฑ ํ•œ๊ตญ์–ด ์™ธ ์–ธ์–ด ์ž…๋ ฅ์€ ์ง€์›ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค

---

## Training Data

### Seed Data

๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ๋ณ„ 5~10๊ฐœ, ์ด ์•ฝ 50~80๊ฐœ์˜ ์‹œ๋“œ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ž‘์—…์œผ๋กœ ์ž‘์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.

| ์นดํ…Œ๊ณ ๋ฆฌ | ๋ชฉํ‘œ ์ƒ˜ํ”Œ ์ˆ˜ | ์„ค๋ช… |
|----------|-------------|------|
| transfer_simple | 150 | ๊ธฐ๋ณธ ์ด์ฒด ("์—„๋งˆํ•œํ…Œ 5๋งŒ์› ๋ณด๋‚ด์ค˜") |
| transfer_complex | 100 | ๋ฉ”๋ชจ ํฌํ•จ, ๋ณตํ•ฉ ์š”์ฒญ |
| confirm_cancel_modify | 80 | ํ™•์ธ/์ทจ์†Œ/์ˆ˜์ • ๋ฉ€ํ‹ฐํ„ด |
| clarify | 80 | ์ •๋ณด ๋ถ€์กฑ ์‹œ ์ž์—ฐ์–ด ๋˜๋ฌผ์Œ |
| query_history | 80 | ๋‚ด์—ญ ์กฐํšŒ (๊ธฐ๊ฐ„/์ˆ˜์ทจ์ธ ํ•„ํ„ฐ) |
| summarize | 60 | ๊ธฐ๊ฐ„๋ณ„ ์ด์ฒด ์š”์•ฝ |
| alias_diversity | 100 | ๋ณ„๋ช… ๋ณ€ํ˜• (์—„๋งˆ/์–ด๋จธ๋‹ˆ/๋ง˜) |
| amount_parsing | 100 | ํ•œ๊ตญ์–ด ๊ธˆ์•ก (์˜ค๋งŒ์›/5๋งŒ/์‚ผ๋ฐฑ๋งŒ) |
| rejection | 50 | ์ด์ฒด ์™ธ ์š”์ฒญ ๊ฑฐ์ ˆ |
| edge_cases | 50 | ์˜คํƒ€, ๋น„๋ฌธ, ํ˜ผํ•ฉ ์š”์ฒญ |

### Data Augmentation

์‹œ๋“œ ๋ฐ์ดํ„ฐ๋ฅผ Claude API๋กœ ์ฆ๊ฐ•ํ•˜์—ฌ 500~1,000๊ฐœ ํ•™์Šต ์ƒ˜ํ”Œ์„ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ์ฆ๊ฐ• ์‹œ ๋‹ค์Œ์„ ๋ณ€ํ˜•ํ•ฉ๋‹ˆ๋‹ค:

- ๋งํˆฌ: ์กด๋Œ“๋ง/๋ฐ˜๋ง/์ค„์ž„๋ง
- ์˜คํƒ€ ๋ฐ ๋น„๋ฌธ
- ๊ธˆ์•ก ํ‘œํ˜„ ๋ฐฉ์‹: ์ˆœํ•œ๊ธ€, ์ˆซ์ž+ํ•œ์ž, ํ˜ผํ•ฉ, ์•„๋ผ๋น„์•„ ์ˆซ์ž
- ๋ณ„๋ช… ๋ณ€ํ˜•

### Data Format

FunctionGemma chat template์„ ์ค€์ˆ˜ํ•ฉ๋‹ˆ๋‹ค:

```jsonl
{
  "messages": [
    {
      "role": "developer",
      "content": "You are a model that can do function calling with the following functions",
      "tool_definitions": [...]
    },
    {
      "role": "user",
      "content": "์—„๋งˆํ•œํ…Œ ์˜ค๋งŒ์› ๋ณด๋‚ด"
    },
    {
      "role": "assistant",
      "content": "",
      "function_calls": [
        {"name": "execute_transfer", "arguments": {"recipient": "์—„๋งˆ", "amount": 50000}}
      ]
    }
  ]
}
```

### Validation

๋ชจ๋“  ๋ฐ์ดํ„ฐ๋Š” ์ž๋™ ๊ฒ€์ฆ์„ ๊ฑฐ์นฉ๋‹ˆ๋‹ค:

- JSON schema ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ
- function name์ด ์ •์˜๋œ 4๊ฐœ ์ค‘ ํ•˜๋‚˜์ธ์ง€ ํ™•์ธ
- amount๊ฐ€ ์–‘์˜ ์ •์ˆ˜์ธ์ง€ ํ™•์ธ
- ํ•œ๊ตญ์–ด ๊ธˆ์•ก -> ์ˆซ์ž ๋ณ€ํ™˜ ์ •ํ™•์„ฑ spot check

---

## Training Procedure

### Base Model

- **๋ชจ๋ธ**: google/functiongemma-270m-it
- **ํ•™์Šต ๋ฐฉ์‹**: Full fine-tuning (๋ชจ๋ธ์ด ๊ฒฝ๋Ÿ‰์ด๋ฏ€๋กœ LoRA ์—†์ด ์ „์ฒด ํŒŒ๋ผ๋ฏธํ„ฐ ํ•™์Šต)

### Hyperparameters

| ํŒŒ๋ผ๋ฏธํ„ฐ | ๊ฐ’ |
|----------|-----|
| Epochs | 5 |
| Batch Size (per device) | 8 |
| Learning Rate | 5e-5 |
| LR Scheduler | cosine |
| Warmup Ratio | 0.1 |
| Weight Decay | 0.01 |
| Max Sequence Length | 2048 |
| Precision | bfloat16 |
| Eval Strategy | epoch |
| Save Strategy | epoch |
| Metric for Best Model | eval_loss |

### Training Environment

- **ํ•˜๋“œ์›จ์–ด**: WSL (RAM 128GB / RTX 3070)
- **์†Œํ”„ํŠธ์›จ์–ด**: HuggingFace Transformers + TRL (SFTTrainer)

### Quantization

```bash
# Fine-tuned ๋ชจ๋ธ -> ONNX ๋ณ€ํ™˜ + INT8 ์–‘์žํ™”
python ml/scripts/convert_onnx.py
```

- ONNX ๋ณ€ํ™˜: optimum (`optimum.exporters.onnx`)
- INT8 ๋™์  ์–‘์žํ™”: onnxruntime (`quantize_dynamic`, `QuantType.QInt8`)
- ์ตœ์ข… ๋ชจ๋ธ ํฌ๊ธฐ: **418MB** (ONNX INT8)
- ์ฐธ๊ณ : INT4๋Š” onnxruntime๊ณผ Gemma weight layout ๋น„ํ˜ธํ™˜์œผ๋กœ INT8 ์‚ฌ์šฉ

---

## Evaluation Results

### Base Model vs Fine-tuned Model ๋น„๊ต

์‹œ๋“œ ๋ฐ์ดํ„ฐ 50๊ฐœ ๊ธฐ์ค€ (ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌํ•จ๋˜์ง€ ์•Š์€ ์›๋ณธ ์‹œ๋“œ)

| ๋ฉ”ํŠธ๋ฆญ | Base (functiongemma-270m-it) | Fine-tuned | ๊ฐœ์„ ์œจ |
|--------|------------------------------|------------|--------|
| Intent Accuracy | 0.0% | 88.9% | +88.9%p |
| JSON Validity | 10.0% | 100.0% | +90.0%p |
| Amount Parsing | 0.0% | 100.0% | +100.0%p |
| Argument F1 (macro) | 0.0% | 57.5% | +57.5%p |
| Rejection Accuracy | 100.0% | 100.0% | +0.0%p |

#### Argument F1 ์„ธ๋ถ€

| ํ•„๋“œ | F1 |
|------|-----|
| recipient | 83.8% |
| amount | 86.1% |
| memo | 75.0% |
| period | 100.0% |
| action | 0.0% |
| new_amount | 0.0% |

> `action`/`new_amount`๋Š” `confirm_transfer` ํ•จ์ˆ˜ ์ „์šฉ ์ธ์ž๋กœ, ์‹œ๋“œ ๋ฐ์ดํ„ฐ ๋‚ด ํ•ด๋‹น ์˜ˆ์ œ ๋ถ€์กฑ์ด ์›์ธ์ž…๋‹ˆ๋‹ค.

---

## Limitations

- **๋„๋ฉ”์ธ ์ œํ•œ**: ์ด์ฒด ๊ด€๋ จ ๋ช…๋ น๋งŒ ์ฒ˜๋ฆฌ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. ๊ทธ ์™ธ ๊ธˆ์œต ์—…๋ฌด(๋Œ€์ถœ, ํˆฌ์ž ๋“ฑ)๋‚˜ ์ผ๋ฐ˜ ๋Œ€ํ™”์—๋Š” ์ ํ•ฉํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
- **์–ธ์–ด ์ œํ•œ**: ํ•œ๊ตญ์–ด ์ž…๋ ฅ๋งŒ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.
- **๋ธŒ๋ผ์šฐ์ € ์ œํ•œ**: WebGPU ์ง€์› ๋ธŒ๋ผ์šฐ์ €(Chrome 113+, Edge 113+)์—์„œ๋งŒ ์ •์ƒ ๋™์ž‘ํ•ฉ๋‹ˆ๋‹ค.
- **๋ฐ๋ชจ ์ „์šฉ**: Mock Banking Engine์œผ๋กœ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋งŒ ์ˆ˜ํ–‰ํ•˜๋ฉฐ, ์‹ค์ œ ๊ธˆ์œต ๊ฑฐ๋ž˜๋ฅผ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
- **ํ•™์Šต ๋ฐ์ดํ„ฐ ํŽธํ–ฅ**: ์‹œ๋“œ ๋ฐ์ดํ„ฐ์™€ Claude API ์ฆ๊ฐ• ๊ธฐ๋ฐ˜์ด๋ฏ€๋กœ, ์‹ค์ œ ์‚ฌ์šฉ์ž ๋ฐœํ™” ํŒจํ„ด๊ณผ ์ฐจ์ด๊ฐ€ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
- **๋ณตํ•ฉ ๋ช…๋ น ์ œํ•œ**: "์—„๋งˆํ•œํ…Œ 5๋งŒ์›, ์•„๋น ํ•œํ…Œ 3๋งŒ์› ๋ณด๋‚ด์ค˜" ๊ฐ™์€ ๋ณตํ•ฉ ์ด์ฒด ๋ช…๋ น์€ ์ง€์›ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

---

## How to Use

### With Transformers.js (Browser)

```javascript
import { pipeline } from '@xenova/transformers';

// ๋ชจ๋ธ ๋กœ๋“œ (WebGPU ์ž๋™ ๊ฐ์ง€)
const generator = await pipeline(
  'text-generation',
  'your-username/transfer-function-gemma-onnx-int4',
  { device: 'webgpu' }
);

// ์ถ”๋ก 
const messages = [
  {
    role: 'system',
    content: 'You are a model that can do function calling with the following functions: [execute_transfer, query_history, summarize_history, confirm_transfer]'
  },
  {
    role: 'user',
    content: '์—„๋งˆํ•œํ…Œ 5๋งŒ์› ๋ณด๋‚ด์ค˜'
  }
];

const output = await generator(messages, {
  max_new_tokens: 256,
  temperature: 0.1,
});

console.log(output);
// => {"name": "execute_transfer", "arguments": {"recipient": "์—„๋งˆ", "amount": 50000}}
```

### With Python (HuggingFace Transformers)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "your-username/transfer-function-gemma",
    torch_dtype="bfloat16",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(
    "your-username/transfer-function-gemma"
)

messages = [
    {"role": "user", "content": "์—„๋งˆํ•œํ…Œ 5๋งŒ์› ๋ณด๋‚ด์ค˜"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True
).to(model.device)

outputs = model.generate(inputs, max_new_tokens=256, temperature=0.1)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)
```

---

## Citation

```bibtex
@misc{transfer-function-gemma-2026,
  title={TransferFunctionGemma: On-Device Korean Banking Function Calling},
  author={Kimin Ryu},
  year={2026},
  url={https://github.com/your-username/TransferFunctionGemma}
}
```

---

## Acknowledgments

- [Google Gemma](https://ai.google.dev/gemma) -- Base model
- [HuggingFace Transformers](https://huggingface.co/docs/transformers) -- Training framework
- [Transformers.js](https://huggingface.co/docs/transformers.js) -- Browser inference
- [Anthropic Claude](https://anthropic.com) -- Data augmentation