|
--- |
|
license: bsd-3-clause |
|
--- |
|
|
|
# CodeT5+ 110M Embedding Models |
|
|
|
## Model description |
|
|
|
[CodeT5+](https://github.com/salesforce/CodeT5/tree/main/CodeT5+) is a new family of open code large language models |
|
with an encoder-decoder architecture that can flexibly operate in different modes (i.e. _encoder-only_, _decoder-only_, |
|
and _encoder-decoder_) to support a wide range of code understanding and generation tasks. |
|
It is introduced in the paper: |
|
|
|
[CodeT5+: Open Code Large Language Models for Code Understanding and Generation](https://arxiv.org/pdf/2305.07922.pdf) |
|
by [Yue Wang](https://yuewang-cuhk.github.io/)\*, [Hung Le](https://sites.google.com/view/henryle2018/home?pli=1)\*, [Akhilesh Deepak Gotmare](https://akhileshgotmare.github.io/), [Nghi D.Q. Bui](https://bdqnghi.github.io/), [Junnan Li](https://sites.google.com/site/junnanlics), [Steven C.H. Hoi](https://sites.google.com/view/stevenhoi/home) (* |
|
indicates equal contribution). |
|
|
|
Compared to the original CodeT5 family (base: `220M`, large: `770M`), CodeT5+ is pretrained with a diverse set of |
|
pretraining tasks including _span denoising_, _causal language modeling_, _contrastive learning_, and _text-code |
|
matching_ to learn rich representations from both unimodal code data and bimodal code-text data. |
|
Additionally, it employs a simple yet effective _compute-efficient pretraining_ method to initialize the model |
|
components with frozen off-the-shelf LLMs such as [CodeGen](https://github.com/salesforce/CodeGen) to efficiently scale |
|
up the model (i.e. `2B`, `6B`, `16B`), and adopts a "shallow encoder and deep decoder" architecture. |
|
Furthermore, it is instruction-tuned to align with natural language instructions (see our InstructCodeT5+ 16B) |
|
following [Code Alpaca](https://github.com/sahil280114/codealpaca). |
|
|
|
## How to use |
|
|
|
This checkpoint consists of an encoder of CodeT5+ 220M model (pretrained from 2 stages on both unimodal and bimodal) and a projection layer, which can be used to extract code |
|
embeddings of 256 dimension. It can be easily loaded using the `AutoModel` functionality and employs the |
|
same [CodeT5](https://github.com/salesforce/CodeT5) tokenizer. |
|
|
|
```python |
|
from transformers import AutoModel, AutoTokenizer |
|
|
|
checkpoint = "Salesforce/codet5p-110m-embedding" |
|
device = "cuda" # for GPU usage or "cpu" for CPU usage |
|
|
|
tokenizer = AutoTokenizer.from_pretrained(checkpoint, trust_remote_code=True) |
|
model = AutoModel.from_pretrained(checkpoint, trust_remote_code=True).to(device) |
|
|
|
inputs = tokenizer.encode("def print_hello_world():\tprint('Hello World!')", return_tensors="pt").to(device) |
|
embedding = model(inputs)[0] |
|
print(f'Dimension of the embedding: {embedding.size()[0]}, with norm={embedding.norm().item()}') |
|
# Dimension of the embedding: 256, with norm=1.0 |
|
print(embedding) |
|
# tensor([ 0.0185, 0.0229, -0.0315, -0.0307, -0.1421, -0.0575, -0.0275, 0.0501, |
|
# 0.0203, 0.0337, -0.0067, -0.0075, -0.0222, -0.0107, -0.0250, -0.0657, |
|
# 0.1571, -0.0994, -0.0370, 0.0164, -0.0948, 0.0490, -0.0352, 0.0907, |
|
# -0.0198, 0.0130, -0.0921, 0.0209, 0.0651, 0.0319, 0.0299, -0.0173, |
|
# -0.0693, -0.0798, -0.0066, -0.0417, 0.1076, 0.0597, -0.0316, 0.0940, |
|
# -0.0313, 0.0993, 0.0931, -0.0427, 0.0256, 0.0297, -0.0561, -0.0155, |
|
# -0.0496, -0.0697, -0.1011, 0.1178, 0.0283, -0.0571, -0.0635, -0.0222, |
|
# 0.0710, -0.0617, 0.0423, -0.0057, 0.0620, -0.0262, 0.0441, 0.0425, |
|
# -0.0413, -0.0245, 0.0043, 0.0185, 0.0060, -0.1727, -0.1152, 0.0655, |
|
# -0.0235, -0.1465, -0.1359, 0.0022, 0.0177, -0.0176, -0.0361, -0.0750, |
|
# -0.0464, -0.0846, -0.0088, 0.0136, -0.0221, 0.0591, 0.0876, -0.0903, |
|
# 0.0271, -0.1165, -0.0169, -0.0566, 0.1173, -0.0801, 0.0430, 0.0236, |
|
# 0.0060, -0.0778, -0.0570, 0.0102, -0.0172, -0.0051, -0.0891, -0.0620, |
|
# -0.0536, 0.0190, -0.0039, -0.0189, -0.0267, -0.0389, -0.0208, 0.0076, |
|
# -0.0676, 0.0630, -0.0962, 0.0418, -0.0172, -0.0229, -0.0452, 0.0401, |
|
# 0.0270, 0.0677, -0.0111, -0.0089, 0.0175, 0.0703, 0.0714, -0.0068, |
|
# 0.1214, -0.0004, 0.0020, 0.0255, 0.0424, -0.0030, 0.0318, 0.1227, |
|
# 0.0676, -0.0723, 0.0970, 0.0637, -0.0140, -0.0283, -0.0120, 0.0343, |
|
# -0.0890, 0.0680, 0.0514, 0.0513, 0.0627, -0.0284, -0.0479, 0.0068, |
|
# -0.0794, 0.0202, 0.0208, -0.0113, -0.0747, 0.0045, -0.0854, -0.0609, |
|
# -0.0078, 0.1168, 0.0618, -0.0223, -0.0755, 0.0182, -0.0128, 0.1116, |
|
# 0.0240, 0.0342, 0.0119, -0.0235, -0.0150, -0.0228, -0.0568, -0.1528, |
|
# 0.0164, -0.0268, 0.0727, -0.0569, 0.1306, 0.0643, -0.0158, -0.1070, |
|
# -0.0107, -0.0139, -0.0363, 0.0366, -0.0986, -0.0628, -0.0277, 0.0316, |
|
# 0.0363, 0.0038, -0.1092, -0.0679, -0.1398, -0.0648, 0.1711, -0.0666, |
|
# 0.0563, 0.0581, 0.0226, 0.0347, -0.0672, -0.0229, -0.0565, 0.0623, |
|
# 0.1089, -0.0687, -0.0901, -0.0073, 0.0426, 0.0870, -0.0390, -0.0144, |
|
# -0.0166, 0.0262, -0.0310, 0.0467, -0.0164, -0.0700, -0.0602, -0.0720, |
|
# -0.0386, 0.0067, -0.0337, -0.0053, 0.0829, 0.1004, 0.0427, 0.0026, |
|
# -0.0537, 0.0951, 0.0584, -0.0583, -0.0208, 0.0124, 0.0067, 0.0403, |
|
# 0.0091, -0.0044, -0.0036, 0.0524, 0.1103, -0.1511, -0.0479, 0.1709, |
|
# 0.0772, 0.0721, -0.0332, 0.0866, 0.0799, -0.0581, 0.0713, 0.0218], |
|
# device='cuda:0', grad_fn=<SelectBackward0>) |
|
``` |
|
|
|
## Pretraining data |
|
|
|
This checkpoint is trained on the stricter permissive subset of the deduplicated version of |
|
the [github-code dataset](https://huggingface.co/datasets/codeparrot/github-code). |
|
The data is preprocessed by reserving only permissively licensed code ("mit" “apache-2”, “bsd-3-clause”, “bsd-2-clause”, |
|
“cc0-1.0”, “unlicense”, “isc”). |
|
Supported languages (9 in total) are as follows: |
|
`c`, `c++`, `c-sharp`, `go`, `java`, `javascript`, `php`, `python`, `ruby.` |
|
|
|
## Training procedure |
|
|
|
This checkpoint is first trained on the unimodal code data at the first-stage pretraining and then on bimodal text-code |
|
pair data using the proposed mixture of pretraining tasks. |
|
Please refer to the paper for more details. |
|
|
|
## Evaluation results |
|
|
|
We show the zero-shot results of this checkpoint on 6 downstream code retrieval tasks from CodeXGLUE in the following table. |
|
| Ruby | JavaScript | Go | Python | Java | PHP | Overall | |
|
| ----- | ---------- | ----- | ------ | ----- | ----- | ------- | |
|
| 74.51 | 69.07 | 90.69 | 71.55 | 71.82 | 67.72 | 74.23 | |
|
|
|
## BibTeX entry and citation info |
|
|
|
```bibtex |
|
@article{wang2023codet5plus, |
|
title={CodeT5+: Open Code Large Language Models for Code Understanding and Generation}, |
|
author={Wang, Yue and Le, Hung and Gotmare, Akhilesh Deepak and Bui, Nghi D.Q. and Li, Junnan and Hoi, Steven C. H.}, |
|
journal={arXiv preprint}, |
|
year={2023} |
|
} |
|
``` |