from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
from tqdm import tqdm
import json

# QLoRA config for 4-bit quantization
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=False,
)

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="auto",
    token=HF_TOKEN
)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, token=HF_TOKEN)

Input Data Processing

The script reads input data from a JSONL file (elyza-tasks-100-TV_0.jsonl). Each line contains a JSON object with task information:

python

datasets = []
with open("./elyza-tasks-100-TV_0.jsonl", "r") as f:
    item = ""
    for line in f:
        line = line.strip()
        item += line
        if item.endswith("}"):
            datasets.append(json.loads(item))
            item = ""

Generation Process

For each input in the dataset:

Format the prompt with instruction template
Tokenize the input
Generate response using the model
Decode the output
Create result object with task_id and output

python

results = []
for data in tqdm(datasets):
    input = data["input"]
    prompt = f"""### Instruction
    {input}
    ### Response:
    """
    
    tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(
            tokenized_input,
            max_new_tokens=100,
            do_sample=False,
            repetition_penalty=1.2
        )[0]
    output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)
    
    results.append({"task_id": data["task_id"], "input": input, "output": output})

Generation Parameters

max_new_tokens=100: Maximum number of tokens to generate
do_sample=False: Deterministic generation (same output every time)
repetition_penalty=1.2: Penalize repetition in generated text

Output Format

The generated responses are saved in a JSONL file with the following format:

{"task_id": "task_1", "input": "input text", "output": "generated response"}

Required fields:

task_id: Unique identifier for the task
output: Response generated by the model

Optional fields:

input: Input text (can be omitted in submission)

Training Data Format

The training data should be provided in JSONL (JSON Lines) format, where each line represents a single JSON object containing the following fields:

json

{
    "instruction": "Task instruction text",
    "input": "Input text (optional)",
    "output": "Expected output text"
}

Fields Description

instruction: Task instruction that tells the model what to do
input: (Optional) Input text that provides specific context for the instruction
output: Expected output that represents the ideal response

Example

json

{"instruction": "以下の文章を要約してください。", "input": "人工知能（AI）は、人間の知能を模倣し、学習、推論、判断などを行うコンピュータシステムです。近年、機械学習や深層学習の発展により、画像認識、自然言語処理、ゲームなど様々な分野で人間に匹敵する、あるいは人間を超える性能を示しています。", "output": "AIは人間の知能を模倣するコンピュータシステムで、機械学習の発展により多くの分野で高い性能を示している。"}
{"instruction": "次の英文を日本語に翻訳してください。", "input": "Artificial Intelligence is transforming the way we live and work.", "output": "人工知能は私たちの生活と仕事の仕方を変革しています。"}

kmagai/llm-jp-3-13b-finetune-2

Uploaded model

JSONL Output Process

Model Inference Setup