---
license: apache-2.0
language:
- ja
---

![image](x1.png)

This model employs the technique described in ["Chat Vector: A Simple Approach to Equip LLMs with Instruction Following and Model Alignment in New Languages"](https://arxiv.org/abs/2310.04799).

It is based on [stablelm-gamma-7b](https://huggingface.co/stabilityai/japanese-stablelm-base-gamma-7b), a model that has not undergone instruction tuning, which was pre-trained using [mistral-7b-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1).

To extract chat vectors, mistral-7b-v0.1 was "subtracted" from [mistral-7b-instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2).

By applying these extracted chat vectors to the non-instruction-tuned model stablelm-gamma-7b, an effect equivalent to instruction tuning is achieved.


```python
from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda" # the device to load the model onto

model = AutoModelForCausalLM.from_pretrained("kousw/stablelm-gamma-7b-chatvector")
tokenizer = AutoTokenizer.from_pretrained("kousw/stablelm-gamma-7b-chatvector")

messages = [
    {"role": "user", "content": "与えられたことわざの意味を小学生でも分かるように教えてください。"},
    {"role": "assistant", "content": "はい、どんなことわざでもわかりやすく答えます"},
    {"role": "user", "content": "情けは人のためならず"}
]

encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt")

model_inputs = encodeds.to(device)
model.to(device)

generated_ids = model.generate(model_inputs, max_new_tokens=256, do_sample=True)
decoded = tokenizer.batch_decode(generated_ids)
print(decoded[0])
```