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 by Yue Wang*, Hung Le*, Akhilesh Deepak Gotmare, Nghi D.Q. Bui, Junnan Li, Steven C.H. Hoi (* 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 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.

This model can be easily loaded using the T5ForConditionalGeneration functionality and employs the same tokenizer as original CodeT5.

This checkpoint is trained on the stricter permissive subset of the deduplicated version of the github-code dataset. 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.

This checkpoint is first trained on the multilingual unimodal code data at the first-stage pretraining, which includes a diverse set of pretraining tasks including span denoising and two variants of causal language modeling. After that, it is further trained on the Python subset with the causal language modeling objective for another epoch to better adapt for Python code generation. Please refer to the paper for more details.

CodeT5+ models have been comprehensively evaluated on a wide range of code understanding and generation tasks in various settings: zero-shot, finetuning, and instruction-tuning. Specifically, CodeT5+ yields substantial performance gains on many downstream tasks compared to their SoTA baselines, e.g., 8 text-to-code retrieval tasks (+3.2 avg. MRR), 2 line-level code completion tasks (+2.1 avg. Exact Match), and 2 retrieval-augmented code generation tasks (+5.8 avg. BLEU-4). In 2 math programming tasks on MathQA-Python and GSM8K-Python, CodeT5+ models of below billion-parameter sizes significantly outperform many LLMs of up to 137B parameters. Particularly, in the zero-shot text-to-code generation task on HumanEval benchmark, InstructCodeT5+ 16B sets new SoTA results of 35.0% pass@1 and 54.5% pass@10 against other open code LLMs, even surpassing the closed-source OpenAI code-cushman-001 mode Please refer to the paper for more details.

Specifically for this checkpoint, it achieves 15.5% pass@1 on HumanEval in the zero-shot setting, which is comparable to much larger LLMs such as Incoder 6B’s 15.2%, GPT-NeoX 20B’s 15.4%, and PaLM 62B’s 15.9%.

This release is for research purposes only in support of an academic paper. Our models, datasets, and code are not specifically designed or evaluated for all downstream purposes. We strongly recommend users evaluate and address potential concerns related to accuracy, safety, and fairness before deploying this model. We encourage users to consider the common limitations of AI, comply with applicable laws, and leverage best practices when selecting use cases, particularly for high-risk scenarios where errors or misuse could significantly impact people’s lives, rights, or safety. For further guidance on use cases, refer to our AUP and AI AUP.