ML4SE23_G1_WizardCoder-SCoT-1B-V1.0

IN4334 ML4SE

Group1 WizardCoder

This model is the result of the fine-tunign of the WizardCoder-1B-V1.0 model using Structured Chain-of-Though (S-CoT) enhanced instructions. S-CoT is used to enhance a sample of about 1200 entries from the Evol-Instruct 80k dataset. The resulting dataset is then used for the training task. The current WizardCoder model and the new S-CoT fine-tuned one are compared on both versions of HumanEval and MBPP (S-CoT enhanced and not) on the pass@1 metric. The S-CoT enhancement of the evaluation datasets allows to study its effect when used just as a prompting technique, independently of the S-CoT fine-tuning of the model.

Fine-tuning Details

Hyperparameter	WizardCoder-1B-V1.0
Batch size	16
Learning rate	2e-5
Epochs	3
Max length	2048
Warmup step	30
LR scheduler	cosine
Dataset	ML4SE23_G1_EvolInstruct-SCoT-1k

The hardware consisted on a GPU instance rented from DataCrunch with the following specifications:

NVidia RTX A6000 48GB 1A6000.10V
2 GPUs
48GB VRAM per GPU
60 GB RAM
10 CPUs
100GB SSD Storage
Ubuntu 20.04
CUDA 11.6

Results

Results of pass@1(%) on HumanEval and MBPP compared to HumanEval-SCoT and MBPP-SCoT using WizardCoder-1B, WizardCoder-SCoT-1B and WizardCoder-15B.

Dataset	WizardCoder-1B-V1.0	WizardCoder-SCoT-1B-V1.0	WizardCoder-15B-V1.0
HumanEval	23.78	17.68	57.3
HumanEval-SCoT	44.51	27.44	57.3
MBPP	23.4	19.4	51.8
MBPP-SCoT	40	28	45.6