llama-2-13b-FINETUNE3_3.3w-r4-q_k_v_o/README.md

license, datasets

license	datasets
llama2	huangyt/FINETUNE3

Model Card for Model ID

在llama-2-13b上使用huangyt/FINETUNE3資料集進行訓練，總資料筆數約3.3w

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
FINETUNE3_3.3w-r4-q_k_v_o	56.29	54.27	79.42	51.90	39.58
FINETUNE3_3.3w-r8-q_k_v_o	56.53	52.99	79.45	53.53	40.14
FINETUNE3_3.3w-r16-q_k_v_o	56.25	53.24	79.53	54.03	38.20
FINETUNE3_3.3w-r4-gate_up_down	55.79	51.02	79.37	53.36	39.40
FINETUNE3_3.3w-r8-gate_up_down	56.60	53.33	79.43	53.60	40.03
FINETUNE3_3.3w-r16-gate_up_down	56.34	51.88	79.42	54.64	39.44
FINETUNE3_3.3w-r4-q_k_v_o_gate_up_down	56.67	53.07	79.34	54.07	40.19
FINETUNE3_3.3w-r8-q_k_v_o_gate_up_down	56.93	54.61	79.16	53.51	40.46
FINETUNE3_3.3w-r16-q_k_v_o_gate_up_down	57.78	53.92	79.41	54.68	43.09

---

• 評估結果來自HuggingFaceH4/open_llm_leaderboard

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA
FINETUNE3_3.3w-r4-q_k_v_o	58.34	59.04	81.15	53	40.16
FINETUNE3_3.3w-r8-q_k_v_o	58.28	56.06	81.89	55.04	40.12
FINETUNE3_3.3w-r16-q_k_v_o	58.55	59.3	81.2	55.58	38.13
FINETUNE3_3.3w-r4-gate_up_down	57.79	56.4	81.93	53.63	39.23
FINETUNE3_3.3w-r8-gate_up_down	58.17	57.25	81.79	53.96	39.66
FINETUNE3_3.3w-r16-gate_up_down	58.91	58.7	81.89	56.08	38.95
FINETUNE3_3.3w-r4-q_k_v_o_gate_up_down	58.42	57.76	80.78	54.32	40.8
FINETUNE3_3.3w-r8-q_k_v_o_gate_up_down	58.26	57.94	81.19	53.43	40.48
FINETUNE3_3.3w-r16-q_k_v_o_gate_up_down	59.62	59.22	81.52	54.94	42.83

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/FINETUNE3", 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 = "FINETUNE3.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}")