wei123602/llama2-13b-fintune2

Model Card for Model ID

在llama-2-13b上使用huangyt/FINETUNE2資料集進行訓練，採用與platypus相近參數

Fine-Tuning Information

• GPU: RTX4090 (single core / 24564MiB)
• base model: meta-llama/Llama-2-13b-hf
• dataset: huangyt/FINETUNE2 (共約3w筆訓練集)
• peft_type: LoRA
• lora_rank: 16
• lora alpha: 8
• lora dropout: 0.05
• lora_target: gate_proj, up_proj, down_proj
• per_device_train_batch_size: 8
• gradient_accumulation_steps: 8
• learning_rate : 4e-4
• epoch: 1
• batch size: 8
• microbatch size: 4
• warmup steps: 100
• weight decay: cosine
• cutoff length: 2048
• precision: bf16
• group_by_length: True
• load_in_8bit: True

Fine-Tuning Detail

• train_loss: 0.0823
• train_runtime: 02:40:01

Evaluation

• 評估結果來自HuggingFaceH4/open_llm_leaderboard
• 與Llama-2-13b比較4種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
CHIH-HUNG/llama-2-13b-FINETUNE2_3w	58.24	58.62	82.32	54.25	38.17
CHIH-HUNG/llama-2-13b-huangyt_Fintune_1_17w-q_k_v_o_proj	58.49	59.73	81.06	54.53	38.64
CHIH-HUNG/llama-2-13b-FINETUNE2_3w-q_k_v_o_proj	58.21	58.53	82.47	53.9	37.92
CHIH-HUNG/llama-2-13b-FINETUNE2_3w-gate_up_down_proj	58.81	57.42	82.42	55.57	39.19
wei123602/llama2-13b-fintune2	wait	wait	wait	wait	wait