llm-jp-3-13b-it_lora

Model Card for yuhkis/llm-jp-3-13b-it_lora

Model Details

Model Description

This is a LoRA-tuned version of LLM-jp-3-13b, fine-tuned using Unsloth techniques and Hugging Face's TRL library for accelerated training.

• Developed by: Yuhki Shiraishi
• Model type: Instruction-tuned Japanese Language Model
• Language: Japanese
• License: CC-BY-NC-SA
• Finetuned from model: llm-jp/llm-jp-3-13b

Uses

Output Generation and Format

Implementation Details

To generate output in the required JSONL format:

# 必要なライブラリをインストール
%%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

# 必要なライブラリを読み込み
from unsloth import FastLanguageModel
from peft import PeftModel
import torch
import json
from tqdm import tqdm
import re

# ベースとなるモデルと学習したLoRAのアダプタ（Hugging FaceのIDを指定）。
model_id = "llm-jp/llm-jp-3-13b"
adapter_id = "yuhkis/llm-jp-3-13b-it_lora"

# Hugging Face Token を指定。
HF_TOKEN = "" #@param {type:"string"}

# unslothのFastLanguageModelで元のモデルをロード。
dtype = None # Noneにしておけば自動で設定
load_in_4bit = True # 今回は13Bモデルを扱うためTrue

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    dtype=dtype,
    load_in_4bit=load_in_4bit,
    trust_remote_code=True,
)

# 元のモデルにLoRAのアダプタを統合。
model = PeftModel.from_pretrained(model, adapter_id, token = HF_TOKEN)

# タスクとなるデータの読み込み。
# 事前にデータをアップロードしてください。
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 = ""

# モデルを用いてタスクの推論。

# 推論するためにモデルのモードを変更
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})

# 結果をjsonlで保存。
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')

Output Format Specification

Required fields in the JSONL output:

• task_id: Task identifier (integer)
• output: Generated response (string)

Example output format:

{"task_id": 1, "output": "生成された応答1"}
{"task_id": 2, "output": "生成された応答2"}

Recommendations

• Ensure consistency in the input prompt structure to maintain output quality.
• Evaluate generated outputs for accuracy, particularly for critical applications.

Training Details

Training Data

• Dataset: Ichikara Instruction Dataset

Training Procedure

• Training regime: bf16 mixed precision, accelerated using Unsloth and Hugging Face's TRL library
• Optimization: LoRA (Low-Rank Adaptation)

Technical Specifications

Model Architecture

• Base model: LLM-jp-3-13b
• Adaptation method: LoRA
• Training enhancement: Unsloth framework

Citation

BibTeX:

@misc{shiraishi2024llm,
    title={LLM-jp-3-13b-it_lora: Instruction-tuned Japanese Language Model with Accelerated Training},
    author={Yuhki Shiraishi},
    year={2024},
    publisher={Hugging Face},
    howpublished={\url{https://huggingface.co/yuhkis/llm-jp-3-13b-it_lora}}
}

Base Model Citation:

@misc{llm-jp2024,
    title={LLM-jp-3: Large Language Model for Japanese},
    author={LLM-jp Project Team},
    year={2024},
    publisher={Hugging Face},
    howpublished={\url{https://huggingface.co/llm-jp/llm-jp-3-13b}}
}

Training Data Citation:

関根聡, 安藤まや, 後藤美知子, 鈴木久美, 河原大輔, 井之上直也, 乾健太郎.
ichikara-instruction: LLMのための日本語インストラクションデータの構築.
言語処理学会第30回年次大会(2024)

Model Card Contact

Primary Contact:

• Name: Yuhki Shiraishi
• GitHub: @yuhkis

For questions regarding this model, please open an issue in the GitHub repository or contact via HuggingFace discussion forum.

Please include "LLM-jp-3-13b-it_lora" in the subject line of any correspondence.