CHIH-HUNG/llama-2-13b-OpenOrca_20w

Model Card for Model ID

在llama-2-13b上使用open orca前20萬筆資料集進行訓練

Fine-Tuning Information

• GPU: RTX4090 (single core / 24564MiB)
• model: meta-llama/Llama-2-13b-hf
• dataset: Open-Orca/OpenOrca (取前20w筆訓練集)
• peft_type: LoRA
• lora_rank: 8
• lora_target: q_proj, v_proj
• per_device_train_batch_size: 8
• gradient_accumulation_steps: 8
• learning_rate : 5e-5
• epoch: 1
• precision: bf16
• quantization: load_in_4bit

Fine-Tuning Detail

• train_loss: 0.8616
• train_runtime: 29:18:07 (use deepspeed)

Evaluation

• 評估結果來自HuggingFaceH4/open_llm_leaderboard
• 與Llama-2-13b和其他使用Open-Orca的模型比較4種Benchmark
• Benchmark包含ARC、HellaSwag、MMLU、TruthfulQA

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA
meta-llama/Llama-2-13b-hf	56.9	58.11	80.97	54.34	34.17
meta-llama/Llama-2-13b-chat-hf	59.93	59.04	81.94	54.64	44.12
Open-Orca/OpenOrca-Platypus2-13B	64.6	62.8	83.15	59.39	53.08
Open-Orca/OpenOrcaxOpenChat-Preview2-13B	63.81	62.37	82.96	58.68	51.23
circulus/Llama-2-13b-orca-v1	62.91	62.03	82.27	57.71	49.61
CHIH-HUNG/llama-2-13b-OpenOrca_5w	61.2	61.01	82.82	56.09	44.87
CHIH-HUNG/llama-2-13b-open_orca_20w	60.46	59.9	82.51	56.3	43.14

How to convert dataset to json

• 在load_dataset中輸入資料集名稱，並且在take中輸入要取前幾筆資料
• 觀察該資料集的欄位名稱，填入example欄位中(例如system_prompt、question、response)
• 最後指定json檔儲存位置 (json_filename)

import json
from datasets import load_dataset

# 讀取數據集，take可以取得該數據集前n筆資料
dataset = load_dataset("Open-Orca/OpenOrca", split="train", streaming=True).take(200000)

# 提取所需欄位並建立新的字典列表
extracted_data = []
for example in dataset:
    extracted_example = {
        ### open orca
        "system_prompt": example["system_prompt"],
        "question": example["question"],
        "response": example["response"]
    }
    extracted_data.append(extracted_example)

# 指定 JSON 文件名稱
json_filename = "open_orca.json"

# 寫入 JSON 文件
with open(json_filename, "w") as json_file:
    json.dump(extracted_data, json_file, indent=4)

print(f"數據已提取並保存為 {json_filename}")