llama-2-13b-FINETUNE4_3.8w-r4-q_k_v_o/README.md

license, datasets

license	datasets
llama2	huangyt/FINETUNE4

Model Card for Model ID

在llama-2-13b上使用huangyt/FINETUNE4資料集進行訓練，總資料筆數約3.8w

Fine-Tuning Information

• GPU: RTX4090 (single core / 24564MiB)
• model: meta-llama/Llama-2-13b-hf
• dataset: huangyt/FINETUNE3 (共約3.8w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
FINETUNE4_3.8w-r4-q_k_v_o	56.67	52.13	79.38	54.54	40.64
FINETUNE4_3.8w-r8-q_k_v_o	56.84	52.30	79.58	54.50	40.98
FINETUNE4_3.8w-r16-q_k_v_o	57.28	53.92	79.92	55.61	39.65
FINETUNE4_3.8w-r4-gate_up_down	55.93	51.71	79.13	53.24	39.63
FINETUNE4_3.8w-r8-gate_up_down	55.93	51.37	79.29	53.62	39.45
FINETUNE4_3.8w-r16-gate_up_down	56.35	52.56	79.28	55.27	38.31
FINETUNE4_3.8w-r4-q_k_v_o_gate_up_down	56.42	53.92	79.09	53.93	38.74
FINETUNE4_3.8w-r8-q_k_v_o_gate_up_down	56.11	51.02	79.24	53.11	41.08
FINETUNE4_3.8w-r16-q_k_v_o_gate_up_down	56.83	53.67	79.49	54.79	39.36

---

• 評估結果來自HuggingFaceH4/open_llm_leaderboard

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA
FINETUNE4_3.8w-r4-q_k_v_o	57.98	54.78	81.4	54.73	41.02
FINETUNE4_3.8w-r8-q_k_v_o	58.96	57.68	81.91	54.95	41.31
FINETUNE4_3.8w-r16-q_k_v_o	58.46	56.23	81.98	55.87	39.76
FINETUNE4_3.8w-r4-gate_up_down	57.94	55.8	81.74	55.09	39.12
FINETUNE4_3.8w-r8-gate_up_down	57.85	54.35	82.13	55.33	39.6
FINETUNE4_3.8w-r16-gate_up_down	57.93	55.03	81.97	56.64	38.07
FINETUNE4_3.8w-r4-q_k_v_o_gate_up_down	58.04	56.31	81.43	55.3	39.11
FINETUNE4_3.8w-r8-q_k_v_o_gate_up_down	58.16	55.97	81.53	54.42	40.72
FINETUNE4_3.8w-r16-q_k_v_o_gate_up_down	58.61	57.25	81.49	55.9	39.79

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/FINETUNE4", 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 = "FINETUNE4.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}")