Tanuki 8B Dpo V1.0

🚀 Tanuki-8B-dpo-v1.0

Tanuki-8B-dpo-v1.0 is a large - scale language model optimized for dialogue, developed by volunteers. It provides various quantized models and has been evaluated in multiple benchmarks.

🚀 Quick Start

To use Tanuki-8B-dpo-v1.0, you can follow the steps in the code example below.

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model = AutoModelForCausalLM.from_pretrained("weblab-GENIAC/Tanuki-8B-dpo-v1.0", device_map="auto", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("weblab-GENIAC/Tanuki-8B-dpo-v1.0")
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

messages = [
    {"role": "system", "content": "以下は、タスクを説明する指示です。要求を適切に満たす応答を書きなさい。"},
    {"role": "user", "content": "たぬきに純粋理性批判は理解できますか？"}
]

input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
output_ids = model.generate(input_ids,
                            max_new_tokens=1024,
                            temperature=0.5,
                            streamer=streamer)

✨ Features

• Model Overview: Tanuki-8B is a large - scale language model with approximately 8B parameters, pre - trained from scratch on about 1.3T tokens. Tanuki-8x8B-dpo-v1.0 is fine - tuned for dialogue using SFT and DPO. For more detailed information, please refer to the blog post.
• Quantized Models:
  • AWQ 4bit Quantization
  • GPTQ 4bit Quantization
  • GPTQ 8bit Quantization
  • GGUF Quantization. Note: The GGUF version may have performance degradation and is not recommended.
• Prompt Format: Tanuki-8B-dpo-v1.0 uses the Japanese Alpaca prompt format.
  • Single - turn:

<s>以下は、タスクを説明する指示です。要求を適切に満たす応答を書きなさい。

### 指示:
たぬきに純粋理性批判は理解できますか？

### 応答:

- **Multi - turn**:

<s>以下は、タスクを説明する指示です。要求を適切に満たす応答を書きなさい。

### 指示:
{1ターン目の入力}

### 応答:
{1ターン目の応答}</s>

### 指示:
{2ターン目の入力}

### 応答:

It is recommended to use the default system prompt "以下は、タスクを説明する指示です。要求を適切に満たす応答を書きなさい。" as the model has not been trained on other system prompts. Please describe the task details in the user prompt.

📚 Documentation

Benchmarks

Human Evaluation

A system mimicking Chatbot Arena was created, and a manual blind test was conducted. (For details, see here). All evaluation data (about 2000 cases) is publicly available.

Japanese MT - Bench

Evaluation by GPT - 4 (gpt - 4 - 0613, scores of - 1 are excluded when calculating the average score)

🔧 Technical Details

The development of Tanuki-8B-dpo-v1.0 was carried out by volunteers (including corporate employees, students, researchers, etc.) gathered through a public call under the GENIAC Matsuo Lab LLM Development Project.

📄 License

This project is licensed under the Apache - 2.0 license.