metadata

language:
  - code
license: apache-2.0
tags:
  - code
  - gpt2
  - generation
datasets:
  - codeparrot/github-code-clean
  - openai_humaneval
metrics:
  - evaluate-metric/code_eval

CodeParrot-Multi 🦜 (small)

CodeParrot-Multi 🦜 is a GPT-2 model (110M parameters) trained to generate code in 9 programming languages: "Java", "JavaScript", "PHP", "Python", "C#", "C++", "GO", "Ruby" and "TypeScript".

Usage

You can load the CodeParrot-Multi model and tokenizer directly in transformers:

from transformers import AutoTokenizer, AutoModelWithLMHead
  
tokenizer = AutoTokenizer.from_pretrained("codeparrot/codeparrot-small-multi")
model = AutoModelWithLMHead.from_pretrained("codeparrot/codeparrot-small-multi")

inputs = tokenizer("def hello_world():", return_tensors="pt")
outputs = model(**inputs)

or with a pipeline:

from transformers import pipeline

pipe = pipeline("text-generation", model="codeparrot/codeparrot-small-multi")
outputs = pipe("def hello_world():")

Training

The model was trained on the small Github code dataset after near deduplication, a subset of Github code dataset with the following settings:

Config	Value
Batch size	192
Context size	1024
Training steps	300'000
Gradient accumulation	2
Gradient checkpointing	False
Learning rate	5e-4
Weight decay	0.1
Warmup steps	2000
Schedule	Cosine

The training was executed on 16 x A100 (40GB) GPUs. This setting amounts to roughly 58 billion tokens.

Performance

We evaluated the model on OpenAI's HumanEval benchmark which consists of programming challenges:

Metric	Value
pass@1	--%
pass@10	--%
pass@100	--%

The pass@k metric tells the probability that at least one out of k generations passes the tests.

Resources

Code: repository