sysong1/dapt-kogpt-orca-sum-adapter

Model Card for Model ID

This repo contains a low-rank adapter for domain-adapted KoGPT fit on the Flan Collection Dataset and a small supervised tuning dataset for summarization.

How to Get Started with the Model

import json
from random import randrange
import torch

from peft import LoraConfig, get_peft_model
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

model1 = AutoModelForCausalLM.from_pretrained(
    "sysong11/dapt-kogpt", torch_dtype="auto", device_map="auto"
)


lora_path = "sysong11/dapt-kogpt-orca-sum-adapter"
model2 = PeftModel.from_pretrained(model1, lora_path, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(lora_path)


test_data = []
with open("./datasets/test.json", "rb") as f:
    for line in f:
        test_data.append(json.loads(line))


prompt_template = """\
<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant"""

msg = "Q:다음 문서를 요약 하세요, Context:{context}"

ix = randrange(len(test_data))
print(ix)
datapoint = test_data[ix]
ref = test_data[ix]["summary_text"]
system_prompt = "You are an AI assistant. User will you give you a task. Your goal is to complete the task as faithfully as you can."
tokens = tokenizer.encode(
    prompt_template.format(
        system_prompt=system_prompt,
        prompt=msg.format(context=datapoint["original_text"]),
    ),
    return_tensors="pt",
).to(device="cuda", non_blocking=True)

gen_tokens = model2.generate(
    input_ids=tokens,
    do_sample=False,
    temperature=0.5,
    max_length=1024,
    pad_token_id=63999,
    eos_token_id=63999,
)
inputs = tokenizer.batch_decode([gen_tokens[0][: tokens[0].shape[0]]])[0]
generated = tokenizer.batch_decode([gen_tokens[0][tokens[0].shape[0] :]])[0].replace(
    "<|im_end|>", ""
)
print(inputs)
print("generated:")
print(generated)

Framework versions

• PEFT 0.7.1