|
|
--- |
|
|
base_model: llm-jp/llm-jp-3-13b |
|
|
tags: |
|
|
- text-generation-inference |
|
|
- transformers |
|
|
- unsloth |
|
|
- llama |
|
|
- trl |
|
|
license: cc-by-4.0 |
|
|
language: |
|
|
- en |
|
|
- ja |
|
|
--- |
|
|
|
|
|
# LLM2024提出モデル |
|
|
|
|
|
# 実行方法 |
|
|
|
|
|
## 1.githubもしくはHuggingFaceのリポジトリより、以下のファイルをダウンロードしてください。 |
|
|
|
|
|
LLM2024_Model_Inference_fewshot2_20241217-L4.ipynb |
|
|
|
|
|
github: https://github.com/h-godai/llm/blob/main/LLM2024_env/LLM2024_Model_Inference_fewshot2_20241217-L4.ipynb |
|
|
|
|
|
## 2."ELYZA-task-100-TV.jsonl"をカレントディレクトリ以下にコピーしてください |
|
|
|
|
|
## 3.「すべて実行」で、必要なファイルのインストールとダウンロードが行われ、カレントディレクトリに評価用のjsonlファイルが出力されます。 |
|
|
|
|
|
# 学習に使用したデータセット |
|
|
|
|
|
tengentoppa |
|
|
https://huggingface.co/datasets/DeL-TaiseiOzaki/Tengentoppa-sft-v1.0 |
|
|
|
|
|
以上 |
|
|
|
|
|
以下、参考までに上記プログラムのコードを貼っておきます。 |
|
|
``` |
|
|
# -*- coding: utf-8 -*- |
|
|
#"""LLM2024_Model_Inference_fewshot2_20241217-L4.ipynb# |
|
|
# |
|
|
#Automatically generated by Colab. |
|
|
# |
|
|
#Original file is located at |
|
|
# https://colab.research.google.com/drive/1kEddvU3WIhShASF95yiua6iOpLDL04jV |
|
|
# |
|
|
#評価準備 |
|
|
# |
|
|
### インストール |
|
|
# |
|
|
#* pip installは失敗することがあるので、失敗したらもう一度実行してください |
|
|
# |
|
|
#* githubから必要なファイルをコピーしてきます。gitが使える環境が必須です。 |
|
|
#""" |
|
|
# |
|
|
# Commented out IPython magic to ensure Python compatibility. |
|
|
# %%capture |
|
|
# # 1. ライブラリのインストール |
|
|
# # 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 |
|
|
|
|
|
!pip install openai |
|
|
!pip install unsloth-zoo # Install or update unsloth-zoo package |
|
|
!pip install --upgrade --no-cache-dir "unsloth-zoo @ git+https://github.com/unslothai/unsloth-zoo.git" |
|
|
|
|
|
!pip install --upgrade openai |
|
|
#!pip install --upgrade transformers[olmo2] |
|
|
|
|
|
!pip install openai #==1.55.3 # This line installs openai==1.55.3 |
|
|
!pip install --upgrade transformers |
|
|
!pip install --upgrade trl |
|
|
|
|
|
# 推論＆評価用ライブラリ h.godai |
|
|
!pip install --upgrade git+https://github.com/h-godai/llm |
|
|
|
|
|
# 必要なファイルをgithubから持ってきます。環境構築した直後に一度だけ実行してください。 |
|
|
|
|
|
!git clone https://github.com/h-godai/llm.git godai_temp |
|
|
!cp -rv "godai_temp/LLM2024_env/" . |
|
|
!rm -r godai_temp |
|
|
|
|
|
#"""２回目以降はここから実行してください。""" |
|
|
|
|
|
from google.colab import userdata |
|
|
|
|
|
# 各APIキーの取得 |
|
|
OPENAI_API_KEY = userdata.get('OPENAI_API_KEY') |
|
|
GEMINI_API_KEY = userdata.get('GEMINI_API_KEY') |
|
|
HF_TOKEN = userdata.get('HF_TOKEN') |
|
|
|
|
|
# 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 |
|
|
|
|
|
import inference_evaluator.InferenceEvaluator as FireStorm |
|
|
|
|
|
DataDir = "./LLM2024_env" |
|
|
|
|
|
# モデルのロード |
|
|
from huggingface_hub import login |
|
|
from peft import LoraConfig |
|
|
|
|
|
adapter_id = None |
|
|
dpo_adapter_id = None |
|
|
model_id = "llm-jp/llm-jp-3-13b" |
|
|
adapter_id = "h-godai/llm-jp-3-13b-ft8-cm25k-dpo1.5x2_bad28_3.44_fs2_LoRA" # Release 15,19 |
|
|
|
|
|
model_name = model_id if adapter_id is None else adapter_id |
|
|
|
|
|
|
|
|
# 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 |
|
|
|
|
|
if adapter_id: |
|
|
# LoRAと元のモデルを一緒にロード |
|
|
model, tokenizer = FastLanguageModel.from_pretrained( |
|
|
model_name=adapter_id, |
|
|
dtype=dtype, |
|
|
load_in_4bit=load_in_4bit, |
|
|
trust_remote_code=True, |
|
|
) |
|
|
else: |
|
|
# 元モデルのみロード |
|
|
model, tokenizer = FastLanguageModel.from_pretrained( |
|
|
model_name=model_id, |
|
|
dtype=dtype, |
|
|
load_in_4bit=load_in_4bit, |
|
|
trust_remote_code=True, |
|
|
) |
|
|
|
|
|
model = FastLanguageModel.for_inference(model) # This line is added to fix the error |
|
|
|
|
|
import json |
|
|
|
|
|
class FewShotGenerator: |
|
|
|
|
|
def __init__(self, fewshot_prompt = None): |
|
|
self.fewshot_prompt = fewshot_prompt |
|
|
|
|
|
def save(self, path): |
|
|
with open(path, 'w', encoding='utf-8') as f: |
|
|
json.dump(self.fewshot_prompt, f, indent=2, ensure_ascii=False) |
|
|
|
|
|
def load(self, path): |
|
|
with open(path, 'r', encoding='utf-8') as f: |
|
|
self.fewshot_prompt = json.load(f) |
|
|
|
|
|
# [{ "keywords": [<keyword[:Group]>,...], "fewshot1": <fewshot>, "endshot": <tailshot> }] |
|
|
# Group指定がある場合、同一グループのいずれかがマッチすればOK |
|
|
# それ以外は全マッチが必要 |
|
|
# |
|
|
def input_prompt_hook(self, eval, prompt): |
|
|
for fsp in self.fewshot_prompt: |
|
|
kwlen = len(fsp["keywords"]) |
|
|
ok = True |
|
|
group = {} |
|
|
for keyword in fsp["keywords"]: |
|
|
if ':' in keyword: |
|
|
# group付、group内のいずれかでOK |
|
|
words = keyword.split(':') |
|
|
keyword = words[0] |
|
|
gr = words[1] |
|
|
hit = keyword in prompt |
|
|
if gr not in group: |
|
|
group[gr] = 0 |
|
|
if hit: |
|
|
group[gr] += 1 |
|
|
else: |
|
|
# groupなし。全て一致する必要あり |
|
|
if keyword not in prompt: |
|
|
ok = False; # 一つでもなければNG |
|
|
pass |
|
|
pass |
|
|
pass # for keyword in |
|
|
if ok: |
|
|
# グループに0があればNG |
|
|
for gr in group: |
|
|
if group[gr] == 0: |
|
|
ok = False; |
|
|
break |
|
|
pass |
|
|
pass |
|
|
pass |
|
|
if ok and fsp["fewshot1"] is not None: |
|
|
if 'endshot' in fsp: |
|
|
return f"{fsp['fewshot1']}\n### 指示\n{prompt}\n\n{fsp['endshot']}\n### 回答\n" |
|
|
else: |
|
|
return f"{fsp['fewshot1']}\n### 指示\n{prompt}\n\n### 回答\n" |
|
|
pass # for fsp in |
|
|
# ない場合はデフォルト |
|
|
return f"{eval.prefix_prompt_}\n### 指示\n{prompt}\n\n### 回答\n" |
|
|
pass # def |
|
|
pass # class |
|
|
|
|
|
fsg = FewShotGenerator() |
|
|
fsg.load(f"{DataDir}/fewshot_prompt.json") |
|
|
|
|
|
|
|
|
# ここら↓は、LLM2024提出用のJsonl出力 |
|
|
|
|
|
# データセットの読み込み。 |
|
|
# omnicampusの開発環境では、左にタスクのjsonlをドラッグアンドドロップしてから実行。 |
|
|
datasets = [] |
|
|
with open(f"./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 = "" |
|
|
|
|
|
# 推論開始 |
|
|
|
|
|
import inference_evaluator.InferenceEvaluator as FireStorm |
|
|
|
|
|
FastLanguageModel.for_inference(model) |
|
|
|
|
|
evaltask = FireStorm.InferenceEvaluator(model, None) # inference only |
|
|
evaltask.tokenizer_ = tokenizer |
|
|
evaltask.max_tokens_ = 1408 # 1024 # max_tokens |
|
|
evaltask.temperature_ = 0.2 # temperature |
|
|
evaltask.repetition_penalty_ = 1.2 # repetition_penalty |
|
|
evaltask.do_sample_ = evaltask.temperature_ > 0 |
|
|
evaltask.top_p = 0.9 # top_p |
|
|
|
|
|
evaltask.prefix_prompt_ = "" # "以下の指示に厳密に従って、正確に回答してください。\n" |
|
|
evaltask.suffix_prompt_ = "" |
|
|
evaltask.input_prompt_hook_ = fsg.input_prompt_hook # few shot hook |
|
|
|
|
|
results = [] |
|
|
for data in datasets: |
|
|
print(f"TaskId: {evaltask.eval_count_}") |
|
|
input = data["input"] |
|
|
output = evaltask.inference(input) |
|
|
print(output) |
|
|
results.append({"task_id": data["task_id"], "input": input, "output": output}) |
|
|
evaltask.eval_count_ += 1 |
|
|
|
|
|
# こちらで生成されたjsolを提出してください。 |
|
|
import re |
|
|
import datetime |
|
|
import pytz |
|
|
now = datetime.datetime.now(pytz.timezone("Asia/Tokyo")).strftime('%Y%m%d_%H%M%S') |
|
|
with open(f"./outputs-{now}.jsonl", 'w', encoding='utf-8') as f: |
|
|
for result in results: |
|
|
print(result) |
|
|
json.dump(result, f, ensure_ascii=False) |
|
|
f.write('\n') |
|
|
``` |
|
|
|
|
|
|
|
|
# Uploaded model |
|
|
|
|
|
- **Developed by:** h-godai |
|
|
- **License:** apache-2.0 |
|
|
- **Finetuned from model :** llm-jp/llm-jp-3-13b |
|
|
|
|
|
This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library. |
|
|
|
|
|
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth) |
