sudominoru/llm-jp-3-13b-it_lora

Uploaded model

• Developed by: sudominoru
• License: apache-2.0
• Finetuned from model : llm-jp/llm-jp-3-13b

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

How to Use the LoRA Adapter

This repository contains a LoRA adapter fine-tuned for efficient inference or further fine-tuning with the base model llm-jp/llm-jp-3-13b.

Installation

Ensure you have the required libraries installed:

pip install unsloth
pip install peft

Loading the Adapter

from unsloth import FastLanguageModel
from peft import PeftModel
import torch

# Hugging Face Token
HF_TOKEN = ""

model_id = "llm-jp/llm-jp-3-13b"
adapter_id = "sudominoru/llm-jp-3-13b-it_lora"

# Load the base model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    trust_remote_code=True,
)

# Add the LoRA adapter
model = PeftModel.from_pretrained(model, adapter_id, token = HF_TOKEN)

Running Inference

# Prepare input data
datasets = [
    {"task_id": 1, "input": "Explain the importance of clean energy."},
    {"task_id": 2, "input": "Translate 'How are you?' to Japanese."},
]

FastLanguageModel.for_inference(model)

results = []
for dt in tqdm(datasets):
    input_text = dt["input"]
    prompt = f"""### 指示\n{input_text}\n### 回答\n"""
    
    # Tokenize input
    inputs = tokenizer([prompt], return_tensors="pt").to(model.device)

    # Generate output
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        use_cache=True,
        do_sample=False,
        repetition_penalty=1.2
    )

    # Decode prediction
    prediction = tokenizer.decode(outputs[0], skip_special_tokens=True).split('\n### 回答')[-1]
    
    results.append({"task_id": dt["task_id"], "input": input_text, "output": prediction})

# Print results
for result in results:
    print(f"Task ID: {result['task_id']}\nInput: {result['input']}\nOutput: {result['output']}\n")