gpt2 code swift mobile generation

iva-codeint-swift-small GPT-2 is (small version - 239.4M parameters) trained from scratch to obtain results in the text-to-code task tailored for Swift language used in native mobile development (iOS).

Usage

from transformers import pipeline

pipe = pipeline("text-generation", model="mvasiliniuc/iva-codeint-swift-small")
outputs = pipe("func triggerNSNotification")

Inference

API_URL = "https://api-inference.huggingface.co/models/mvasiliniuc/iva-codeint-swift-small"
headers = {"Authorization": "Bearer <key>"}
def query(payload):
	response = requests.post(API_URL, headers=headers, json=payload)
	return response.json()

output = query({
"inputs": """
/* 
A function that gets the current device operating system.
*/
"""
})
pprint.pprint(output, compact=True)

Training

Config Value
seq length 1024
weight decay 0.1
learning rate 0.0005
max eval steps -1
shuffle buffer 10000
max train steps 150000
mixed precision fp16
num warmup steps 2000
train batch size 5
valid batch size 5
lr scheduler type cosine
save checkpoint steps 15000
gradient checkpointing false
gradient accumulation steps 1

Resources

Resources used for research: