Developed by :
- Changgil Song
Model Number:
- k2s3_test_24001
Base Model :
Training Data
- The model was trained on a diverse dataset comprising approximately 800 million tokens, including the Standard Korean Dictionary, KULLM training data from Korea University, dissertation abstracts from master's and doctoral theses, and Korean language samples from AI Hub.
- μ΄ λͺ¨λΈμ νμ€λκ΅μ΄μ¬μ , κ³ λ €λ KULLMμ νλ ¨ λ°μ΄ν°, μλ°μ¬νμμ μμ§μ 보 λ Όλ¬Έμ΄λ‘, ai_hubμ νκ΅μ΄ λ°μ΄ν° μνλ€μ ν¬ν¨νμ¬ μ½ 8μ΅ κ°μ ν ν°μΌλ‘ ꡬμ±λ λ€μν λ°μ΄ν°μ μμ νλ ¨λμμ΅λλ€.
Training Method
- This model was fine-tuned on the "meta-llama/Llama-2-13b-chat-hf" base model using PEFT (Parameter-Efficient Fine-Tuning) LoRA (Low-Rank Adaptation) techniques.
- μ΄ λͺ¨λΈμ "meta-llama/Llama-2-13b-chat-hf" κΈ°λ° λͺ¨λΈμ PEFT LoRAλ₯Ό μ¬μ©νμ¬ λ―ΈμΈμ‘°μ λμμ΅λλ€.
Hardware and Software
- Hardware: Utilized two A100 (80G*2EA) GPUs for training.
- Training Factors: This model was fine-tuned using PEFT LoRA with the HuggingFace SFTtrainer and applied fsdp. Key parameters included LoRA r = 8, LoRA alpha = 16, trained for 2 epochs, batch size of 1, and gradient accumulation of 32.
- μ΄ λͺ¨λΈμ PEFT LoRAλ₯Ό μ¬μ©νμ¬ HuggingFace SFTtrainerμ fsdpλ₯Ό μ μ©νμ¬ λ―ΈμΈμ‘°μ λμμ΅λλ€. μ£Όμ νλΌλ―Έν°λ‘λ LoRA r = 8, LoRA alpha = 16, 2 μν νλ ¨, λ°°μΉ ν¬κΈ° 1, κ·Έλ¦¬κ³ κ·ΈλΌλμΈνΈ λμ 32λ₯Ό ν¬ν¨ν©λλ€.
Caution
- For fine-tuning this model, it is advised to consider the specific parameters used during training, such as LoRA r and LoRA alpha values, to ensure compatibility and optimal performance.
- μ΄ λͺ¨λΈμ λ―ΈμΈμ‘°μ ν λλ LoRA r λ° LoRA alpha κ°κ³Ό κ°μ΄ νλ ¨ μ€μ μ¬μ©λ νΉμ νλΌλ―Έν°λ₯Ό κ³ λ €νλ κ²μ΄ μ’μ΅λλ€. μ΄λ νΈνμ± λ° μ΅μ μ μ±λ₯μ 보μ₯νκΈ° μν¨μ λλ€.
Additional Information
- The training leveraged the fsdp (Fully Sharded Data Parallel) feature through the HuggingFace SFTtrainer for efficient memory usage and accelerated training.
- νλ ¨μ HuggingFace SFTtrainerλ₯Ό ν΅ν fsdp κΈ°λ₯μ νμ©νμ¬ λ©λͺ¨λ¦¬ μ¬μ©μ ν¨μ¨μ μΌλ‘ νκ³ νλ ¨ μλλ₯Ό κ°μννμ΅λλ€.
- Downloads last month
- 86