You can deploy this model and start earning money today!

hiroki-rad/
llm-jp-llm-jp-3-13b-128-ft-3000

Name: llm-jp-llm-jp-3-13b-128-ft-3000
Availability: InStock

safetensors

llama

hiroki-rad/elyza_generated_data-3031

llm-jp/llm-jp-3-13b

finetune

Views

No views yet

python

!pip install -U langchain-community langchain-huggingface vllm triton wandb weave langchain-huggingface langchain-chroma datasets --q

# GitHubをclone
!git clone https://github.com/y-hiroki-radiotech/llm-final-task.git
%cd llm-final-task

import os
import random
import numpy as np
import torch
import pandas as pd
from vllm import LLM
from tqdm import tqdm
import json
from datasets import
import time

from custom_few_shot_prompt_template import CustomFewShotPromptTemplate


# JSONLファイルを読み込む
file_path = 'elyza-tasks-100-TV_0.jsonl'
data = pd.read_json(file_path, lines=True)

# example selector用のデータ
df = load_dataset("elyza/ELYZA-tasks-100", split="test")
df = df.to_pandas()
examples = []
for row in df.itertuples():
    examples.append({"input": row.input, "output": row.output})


few_shot = CustomFewShotPromptTemplate(examples)
# few-shot-selector
few_shot_list = []
for row in tqdm(data.itertuples(), desc="生成中"):
    few_shot_list.append(few_shot.format(row.input))

# 一度キャッシュを削除する
if torch.cuda.is_available():
    print("Clearing CUDA cache...")
    torch.cuda.empty_cache()
    print("CUDA cache cleared.")
else:
    print("CUDA is not available on this system.")

# vllmを使う
model_name = "hiroki-rad/llm-jp-llm-jp-3-13b-128-ft-3000"
llm = LLM(model=model_name)

# 2回考えるように推論するクラスをインスタンス化
from two_stage_think import TwoStageThinking


thinking_generator = TwoStageThinking(llm)

# 最終的に1回推論の回答を使うことにした
results = []
for row, few_shot in tqdm(zip(data.itertuples(), few_shot_list), desc="生成中"):
    time.sleep(3)
    first = thinking_generator.generate_complete_response(row, few_shot)
    results.append(first)

# データの格納
jsonl_data = []
for i in range(len(data)):
    task_id = data.iloc[i]["task_id"] # Access task_id using the index
    output = results[i]

    # Create a dictionary for each row
    jsonl_object = {
        "task_id": task_id,
        "output": output
    }
    jsonl_data.append(jsonl_object)

with open("llm-jp-vllm-second-thinking-output.jsonl", "w", encoding="utf-8") as outfile:
  for entry in jsonl_data:
    # Convert task_id to a regular Python integer before dumping
    entry["task_id"] = int(entry["task_id"])
    json.dump(entry, outfile, ensure_ascii=False)
    outfile.write('\n')