suka-nlp/Llama_2_finetune

Model Details

Model Description

• Developed by: Luthfantry Zamzam Muhammad
• Model type: large language model for code generation
• Language(s) (NLP): English
• License: [More Information Needed]
• Finetuned from model: llama2

Model Sources

• Repository: https://github.com/unslothai/unsloth
• Developed by: unsloth

Model parameter

• r = 16,
• target_modules = ["q_proj", "k_proj", "v_proj", "o_proj","gate_proj", "up_proj", "down_proj",],
• lora_alpha = 16,
• lora_dropout = 0,
• bias = "none",
• use_gradient_checkpointing = "unsloth",
• random_state = 3407,
• use_rslora = False,
• loftq_config = None,

Usage and limitations

This model is used to generate code based on commands given by the user. it should be noted that this model can generate many languages because it takes the initial model from llama2. However, after finetuning it is better at generating python code, because currently it is only trained with python code datasets.

How to Get Started with the Model

use link below to use model https://colab.research.google.com/drive/1lBzz5KeZJKXjvivbYvmGarix9Ao6Wxe5?usp=sharing

Training Data

https://huggingface.co/datasets/mlabonne/Evol-Instruct-Python-26k

Training Hyperparameters

• Warmup_step: 5
• lr_scheduler_type: linear
• Learning Rate: 0.0002
• Batch Size: 2
• Weigh_decay: 0.001
• Epoch: 60
• Optimizer: adamw_8bit

Testing Data

https://huggingface.co/datasets/google-research-datasets/mbpp/viewer/full

Testing Document

https://docs.google.com/spreadsheets/d/1hr8R4nixQsDC5cGGENTOLUW1jPCS_lltVRIeOzenBvA/edit?usp=sharing

Results

Berfore finetune

• Accurary : 1,8%
• Consistensy : 0%

After fine tune

• Accuracy : 49%
• Consistency : 100%