RichardErkhov/lemon-mint_-_Llama-VARCO-8B-Instruct-LLaMAfied-awq
Quantization made by Richard Erkhov.
Llama-VARCO-8B-Instruct-LLaMAfied - AWQ
- Model creator: https://huggingface.co/lemon-mint/
- Original model: https://huggingface.co/lemon-mint/Llama-VARCO-8B-Instruct-LLaMAfied/
Original model description: --- language:
- en
- ko license: llama3.1 tags:
- llama-3.1
- ncsoft
- varco base_model:
- meta-llama/Meta-Llama-3.1-8B library_name: transformers ---
Llama-VARCO-8B-Instruct
About the Model
Llama-VARCO-8B-Instruct is a generative model built with Llama, specifically designed to excel in Korean through additional training. The model uses continual pre-training with both Korean and English datasets to enhance its understanding and generation capabilites in Korean, while also maintaining its proficiency in English. It performs supervised fine-tuning (SFT) and direct preference optimization (DPO) in Korean to align with human preferences.
- Developed by: NC Research, Language Model Team
- Languages (NLP): Korean, English
- License: LLAMA 3.1 COMMUNITY LICENSE AGREEMENT
- Base model: meta-llama/Meta-Llama-3.1-8B
Uses
Direct Use
We recommend to use transformers v4.43.0 or later, as advised for Llama-3.1.
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model = AutoModelForCausalLM.from_pretrained(
"NCSOFT/Llama-VARCO-8B-Instruct",
torch_dtype=torch.bfloat16,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("NCSOFT/Llama-VARCO-8B-Instruct")
messages = [
{"role": "system", "content": "You are a helpful assistant Varco. Respond accurately and diligently according to the user's instructions."},
{"role": "user", "content": "안녕하세요."}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
eos_token_id = [
tokenizer.eos_token_id,
tokenizer.convert_tokens_to_ids("<|eot_id|>")
]
outputs = model.generate(
inputs,
eos_token_id=eos_token_id,
max_length=8192
)
print(tokenizer.decode(outputs[0]))Evaluation
LogicKor
We used the LogicKor code to measure performance. For the judge model, we used the officially recommended gpt-4-1106-preview. The score includes only the 0-shot evaluation provided in the default.