llama-2-13b-FINETUNE5_4w-r4-q_k_v_o/README.md

license, datasets

license	datasets
llama2	huangyt/FINETUNE5

Model Card for Model ID

在llama-2-13b上使用huangyt/FINETUNE5資料集進行訓練，總資料筆數約4w

Fine-Tuning Information

• GPU: RTX4090 (single core / 24564MiB)
• model: meta-llama/Llama-2-13b-hf
• dataset: huangyt/FINETUNE3 (共約3.3w筆訓練集)
• peft_type: LoRA
• lora_rank: 16
• lora_target: q_proj, k_proj, v_proj, o_proj
• per_device_train_batch_size: 8
• gradient_accumulation_steps: 8
• learning_rate : 4e-4
• epoch: 1
• precision: bf16
• quantization: load_in_4bit

Fine-Tuning Detail

• train_loss: 0.579
• train_runtime: 4:6:11 (use deepspeed)

Evaluation

• 與Llama-2-13b比較4種Benchmark，包含ARC、HellaSwag、MMLU、TruthfulQA
• 評估結果使用本地所測的分數，並使用load_in_8bit

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA	Time (s)
FINETUNE5_4w-r4-q_k_v_o	56.09	54.35	79.24	54.01	36.75	22095
FINETUNE5_4w-r8-q_k_v_o	57.55	55.38	79.57	54.03	41.21	22127
FINETUNE5_4w-r16-q_k_v_o	57.26	54.35	79.74	52.29	42.68	22153
FINETUNE5_4w-r4-gate_up_down	56.51	52.82	79.13	52.83	41.28	22899
FINETUNE5_4w-r8-gate_up_down	56.10	52.73	79.14	52.56	39.99	22926
FINETUNE5_4w-r16-gate_up_down	56.23	52.39	79.48	53.42	39.62	22963
FINETUNE5_4w-r4-q_k_v_o_gate_up_down	56.06	52.56	79.21	51.67	40.80	24303
FINETUNE5_4w-r8-q_k_v_o_gate_up_down	56.35	51.88	79.42	52.00	42.10	24376
FINETUNE5_4w-r16-q_k_v_o_gate_up_down	56.73	54.18	79.53	52.77	40.46	24439

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("huangyt/FINETUNE5", split="train", streaming=True)

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

# 指定 JSON 文件名稱
json_filename = "FINETUNE5.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}")