baku-10b-chat-v2 / README.md
sudy-super's picture
Update README.md
708defa
metadata
license: apache-2.0
language:
  - ja
  - en

Description

This model is a 10.2 billion parameter model that combines two sets of 24 layers each from CALM2-7B-chat using slerp-merge.

Chat Template

USER: {user_message1}
ASSISTANT: {assistant_message1}<|endoftext|>
USER: {user_message2}
ASSISTANT: {assistant_message2}<|endoftext|>
USER: {user_message3}
ASSISTANT: {assistant_message3}<|endoftext|>

Tutorial

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("sudy-super/baku-10b-chat-v2")
model = AutoModelForCausalLM.from_pretrained("sudy-super/baku-10b-chat-v2", device_map="auto", torch_dtype=torch.bfloat16)

raw_prompt = "仕事の熱意を取り戻すためのアイデアを5つ挙げてください。"
prompt = f"USER:{raw_prompt}\nASSISTANT:"

token_ids = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
with torch.no_grad():
    output_ids = model.generate(
        token_ids.to(model.device),
        max_new_tokens=100,
        do_sample=True,
        temperature=0.8,
        pad_token_id=tokenizer.pad_token_id,
        bos_token_id=tokenizer.bos_token_id,
        eos_token_id=tokenizer.eos_token_id
    )

result = tokenizer.decode(output_ids.tolist()[0])
print(result)