---
language:
  - zh
  - en
license: apache-2.0
base_model: Qwen/Qwen3.5-4B
tags:
  - lora
  - qlora
  - roleplay
  - character-ai
  - taiwanese-mandarin
  - llama-factory
  - gguf
datasets:
  - RX5950XTP/silicon-girlfriend-dataset
---
# Silicon-Based-Girlfriend — QLoRA Adapter

---

基於 **Qwen3.5-4B** 的 QLoRA 微調 Adapter，訓練目標為沉浸式繁體中文角色扮演。本倉庫包含 LoRA Adapter 權重與 GGUF 格式模型。

---

## Model Details / 模型資訊

| 項目               | 內容                                                   |
| ------------------ | ------------------------------------------------------ |
| Base Model         | [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) |
| Fine-tuning Method | QLoRA (4-bit NF4)                                      |
| LoRA Rank          | 32                                                     |
| LoRA Alpha         | 64                                                     |
| LoRA Dropout       | 0.05                                                   |
| LoRA Target        | All linear layers                                      |
| Training Epochs    | 5                                                      |
| Context Length     | 8192 tokens                                            |
| Learning Rate      | 1e-4                                                   |
| LR Scheduler       | Cosine                                                 |
| Optimizer          | paged_adamw_8bit                                       |
| Training Samples   | 985                                                    |
| Train Loss         | 1.108                                                  |
| Eval Loss          | 1.434                                                  |
| Hardware           | NVIDIA RTX A6000 (48GB VRAM)                           |
| Training Time      | ~19 hours                                              |
| Framework          | [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) |
| Chat Template      | `qwen3_5_nothink` (non-thinking mode)                |

---

## Files / 檔案說明

| 檔案                            | 說明                                                 |
| ------------------------------- | ---------------------------------------------------- |
| `adapter_config.json`         | LoRA 設定檔                                          |
| `adapter_model.safetensors`   | LoRA 權重（248 MB）                                  |
| `tokenizer_config.json`       | Tokenizer 設定（含 nothink chat template）           |
| `tokenizer.json`              | Tokenizer                                            |
| `vocab.json` / `merges.txt` | Vocabulary                                           |
| `silicon-gf-q8_0.gguf`        | Q8_0 量化 GGUF（4.2 GB，適用 llama.cpp / LM Studio） |
| `training_loss.png`           | 訓練 Loss 曲線                                       |
| `training_eval_loss.png`      | 評估 Loss 曲線                                       |

---

## Usage / 使用方式

### Option 1: GGUF (Recommended / 推薦)

直接在 **LM Studio** 或 **llama.cpp** 載入 `silicon-gf-q8_0.gguf`，無需額外安裝。

```bash
# llama.cpp
./llama-cli -m silicon-gf-q8_0.gguf -c 8192 --temp 0.8
```

### Option 2: LoRA Adapter with transformers + PEFT

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "Qwen/Qwen3.5-4B"
adapter = "RX5950XTP/silicon-based-girlfriend"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

messages = [
    {"role": "user", "content": "嘿，你在幹嘛？"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.8, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
```

### Option 3: LLaMA-Factory inference

```bash
llamafactory-cli chat \
  --model_name_or_path Qwen/Qwen3.5-4B \
  --adapter_name_or_path RX5950XTP/silicon-based-girlfriend \
  --template qwen3_5_nothink \
  --finetuning_type lora
```

---

## Training Curves / 訓練曲線

![Training Loss](training_loss.png)
![Eval Loss](training_eval_loss.png)

---

## Dataset / 訓練資料集

- **倉庫**：[RX5950XTP/silicon-girlfriend-dataset](https://huggingface.co/datasets/RX5950XTP/silicon-girlfriend-dataset)
- **格式**：ShareGPT（`system` + `conversations` with `from`/`value`）
- **筆數**：985 筆多輪對話
- **語言**：繁體中文（臺灣用語）
- **生成方式**：由 Kimi K2.5 根據角色設定生成

---

## Notes / 注意事項

- 本模型使用 `qwen3_5_nothink` chat template，**預設不啟用思考模式**，回覆會直接輸出角色對話。
- 角色設定包含不良用語與成人主題，請自行評估使用場景。
- 模型以 QLoRA 訓練，推理時需搭配 base model（Qwen3.5-4B）一同載入，或直接使用 GGUF。

---

## License

Apache 2.0（遵循 Qwen3.5-4B 原授權）