llm-jp-3-13b-it_lora

Llama Model with Unsloth and TRL

Overview

This repository contains a fine-tuned Llama model, developed by Ohais, using Unsloth and Huggingface's TRL library. The fine-tuning process achieved a 2x speed improvement compared to traditional methods. The model is derived from llm-jp/llm-jp-3-13b and is distributed under the Apache-2.0 license.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>

Features

• Base Model: llm-jp/llm-jp-3-13b
• Fine-tuned Using: Unsloth and Huggingface's TRL library
• License: Apache-2.0
• Language: English

Inference Guide

Follow the steps below to perform inference with this model.

Step 1: Install Required Libraries

%%capture
!pip install unsloth
!pip uninstall unsloth -y && pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install -U torch
!pip install -U peft

Step 2: Load Required Libraries

from unsloth import FastLanguageModel
from peft import PeftModel
import torch
import json
from tqdm import tqdm
import re

Step 3: Load the Model

Specify the base model and the adapter for LoRA fine-tuning. Replace <adapter_id> and <HF_TOKEN> with appropriate values.

# Base model and LoRA adapter
model_id = "llm-jp/llm-jp-3-13b"
adapter_id = "Ohais/llm-jp-3-13b-it_lora"

# Hugging Face Token
HF_TOKEN = "<your_hf_token>"  # Obtain token from https://huggingface.co/settings/tokens

# Load base model using Unsloth
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    dtype=None,  # Automatically set dtype
    load_in_4bit=True,  # Recommended for large models
    trust_remote_code=True,
)

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

Step 4: Load Dataset

Prepare your dataset in .jsonl format and upload it to your environment.

# Load task data
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 = ""

Step 5: Perform Inference

Set the model to inference mode and generate predictions.

# Set model to inference mode
FastLanguageModel.for_inference(model)

results = []
for dt in tqdm(datasets):
    input = dt["input"]

    prompt = f"""### 指示\n{input}\n### 回答\n"""

    inputs = tokenizer([prompt], return_tensors="pt").to(model.device)

    outputs = model.generate(
        **inputs, max_new_tokens=512, use_cache=True, do_sample=False, repetition_penalty=1.2
    )
    prediction = tokenizer.decode(outputs[0], skip_special_tokens=True).split('\n### 回答')[-1]

    results.append({"task_id": dt["task_id"], "input": input, "output": prediction})

Step 6: Save Results

Save the inference results to a .jsonl file. Replace <adapter_id> with the appropriate identifier.

# Save results to JSONL
json_file_id = re.sub(".*/", "", adapter_id)
with open(f"/content/{json_file_id}_output.jsonl", 'w', encoding='utf-8') as f:
    for result in results:
        json.dump(result, f, ensure_ascii=False)
        f.write('\n')

License

This project is licensed under the Apache-2.0 License. See the LICENSE file for more details.