119 lines
2.9 KiB
Markdown
119 lines
2.9 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
language:
|
|||
|
|
- zh
|
|||
|
|
base_model: Qwen/Qwen3-1.7B
|
|||
|
|
tags:
|
|||
|
|
- conversational
|
|||
|
|
- sft
|
|||
|
|
- coffee
|
|||
|
|
- traditional-chinese
|
|||
|
|
- qwen3
|
|||
|
|
- task-oriented-dialogue
|
|||
|
|
datasets:
|
|||
|
|
- renhehuang/coffee-order-zhtw
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Qwen3-1.7B Coffee Order Assistant (繁體中文)
|
|||
|
|
|
|||
|
|
基於 [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) 進行全參數 SFT 微調的咖啡點餐助理模型。
|
|||
|
|
|
|||
|
|
> 💡 量化版本（INT4，~1.45 GB）：[renhehuang/qwen3-1.7b-coffee-sft-quanto-int4](https://huggingface.co/renhehuang/qwen3-1.7b-coffee-sft-quanto-int4)
|
|||
|
|
|
|||
|
|
## 模型描述
|
|||
|
|
|
|||
|
|
此模型專為繁體中文咖啡點餐場景設計，能夠：
|
|||
|
|
- 理解用戶的點餐意圖
|
|||
|
|
- 處理多輪對話
|
|||
|
|
- 確認訂單細節（飲品種類、冰熱、加濃縮咖啡）
|
|||
|
|
- 處理訂單修改與取消
|
|||
|
|
|
|||
|
|
### 支援的菜單項目
|
|||
|
|
- 美式
|
|||
|
|
- 拿鐵
|
|||
|
|
- 燕麥奶拿鐵
|
|||
|
|
- 鮮奶
|
|||
|
|
|
|||
|
|
### 可選項目
|
|||
|
|
- 冰/熱
|
|||
|
|
- 加一份濃縮咖啡
|
|||
|
|
|
|||
|
|
## 使用方式
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
|
|||
|
|
model_name = "renhehuang/qwen3-1.7b-coffee-sft"
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
|
|||
|
|
|
|||
|
|
messages = [
|
|||
|
|
{"role": "system", "content": "你是一位專業的咖啡點餐助理，負責協助使用者完成點餐。菜單包含：美式、拿鐵、燕麥奶拿鐵、鮮奶。"},
|
|||
|
|
{"role": "user", "content": "我想要一杯冰拿鐵"}
|
|||
|
|
]
|
|||
|
|
|
|||
|
|
input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
|||
|
|
inputs = tokenizer(input_text, return_tensors="pt")
|
|||
|
|
|
|||
|
|
outputs = model.generate(
|
|||
|
|
**inputs,
|
|||
|
|
max_new_tokens=128,
|
|||
|
|
do_sample=True,
|
|||
|
|
temperature=0.7,
|
|||
|
|
top_p=0.9,
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
|||
|
|
print(response)
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## 訓練細節
|
|||
|
|
|
|||
|
|
### 訓練配置
|
|||
|
|
| 參數 | 值 |
|
|||
|
|
|------|-----|
|
|||
|
|
| 基礎模型 | Qwen/Qwen3-1.7B |
|
|||
|
|
| 訓練方式 | Full Parameter SFT |
|
|||
|
|
| 學習率 | 5e-6 |
|
|||
|
|
| Batch Size | 1 |
|
|||
|
|
| Gradient Accumulation | 16 |
|
|||
|
|
| Epochs | 3 |
|
|||
|
|
| Max Length | 512 |
|
|||
|
|
| Optimizer | AdamW |
|
|||
|
|
| LR Scheduler | Cosine |
|
|||
|
|
| Warmup Ratio | 0.1 |
|
|||
|
|
|
|||
|
|
### 訓練資料
|
|||
|
|
- 資料集：[renhehuang/coffee-order-zhtw](https://huggingface.co/datasets/renhehuang/coffee-order-zhtw)
|
|||
|
|
- 資料筆數：約 2900+ 筆多輪對話
|
|||
|
|
- 語言：繁體中文（台灣）
|
|||
|
|
|
|||
|
|
### 硬體
|
|||
|
|
- Apple Silicon (MPS)
|
|||
|
|
|
|||
|
|
## 限制與注意事項
|
|||
|
|
|
|||
|
|
- 此模型僅針對咖啡點餐場景進行訓練，不適用於一般對話
|
|||
|
|
- 菜單項目固定，無法處理菜單外的飲品
|
|||
|
|
- 每杯飲品最多只能加一份濃縮咖啡
|
|||
|
|
- 所有飲品統一為大杯
|
|||
|
|
|
|||
|
|
## 授權
|
|||
|
|
|
|||
|
|
本模型基於 Apache 2.0 授權發布。
|
|||
|
|
|
|||
|
|
## 引用
|
|||
|
|
|
|||
|
|
如果您使用此模型，請引用：
|
|||
|
|
|
|||
|
|
```bibtex
|
|||
|
|
@misc{qwen3-coffee-sft,
|
|||
|
|
author = {Ren-He Huang},
|
|||
|
|
title = {Qwen3-1.7B Coffee Order Assistant},
|
|||
|
|
year = {2025},
|
|||
|
|
publisher = {HuggingFace},
|
|||
|
|
url = {https://huggingface.co/renhehuang/qwen3-1.7b-coffee-sft}
|
|||
|
|
}
|
|||
|
|
```
|