Repository and research release summary for Salesforce CodeT5 and CodeT5+. Main repository: https://github.com/salesforce/CodeT5 (license indicated on the GitHub page: BSD-3-Clause). Primary papers: CodeT5 (EMNLP 2021) available at https://arxiv.org/abs/2109.00859 (introduces identifier-aware encoder–decoder pretraining with identifier masking and bimodal dual generation unified for code understanding and generation) and CodeT5+ (2023) at https://arxiv.org/abs/2305.07922 (describes flexible encoder–decoder large models with mixed pretraining objectives, instruction tuning, and large open checkpoints). Available Hugging Face checkpoints and model cards: Salesforce/codet5-small (https://huggingface.co/Salesforce/codet5-small), Salesforce/codet5-base (https://huggingface.co/Salesforce/codet5-base), Salesforce/codet5-large (https://huggingface.co/Salesforce/codet5-large), Salesforce/codet5-large-ntp-py (https://huggingface.co/Salesforce/codet5-large-ntp-py), and Salesforce/codet5p-220m (CodeT5+ checkpoint) (https://huggingface.co/Salesforce/codet5p-220m). Main capabilities documented across the repo, papers, and model cards include: text-to-code generation (NL -> code), code autocompletion (including full function completion from name/signature), code summarization (code -> NL), code translation (PL -> PL), refinement/repair, defect detection, clone detection, retrieval, and instruction-tuned variants (CodeT5+) for improved zero-shot/instruction behavior. Notable technical details: CodeT5 introduces identifier-aware pretraining and bimodal tasks to improve NL–PL alignment and preserve code semantics. CodeT5+ uses a mixture of pretraining objectives (span denoising, contrastive learning, text-code matching, causal LM) and can initialize from frozen LLM components to scale efficiently; it reports state-of-the-art numbers among open models on many code tasks according to the authors. Salesforce provides a demo (including a VS Code plugin) and invites sharing builds via [email protected]. The repository and model cards include ethical/usage guidance and restrictions (e.g., not to use models to promote violence, hate, environmental destruction, human-rights abuses, or harmful health recommendations) and advise following AI deployment documentation for high-stakes systems. What could not be obtained automatically here: the raw README and LICENSE file contents — automated raw-file fetch attempts triggered GitHub’s abuse-detection or returned 404 when pulling raw files via the scraper used; the GitHub HTML page indicates BSD-3-Clause but raw downloads of specific files failed in this automated fetch. Recommended ways to fetch full repo files: clone the repo locally (git clone https://github.com/salesforce/CodeT5.git) or browse the README and LICENSE via the GitHub web UI, or use the GitHub API with authentication (token) to avoid abuse-detection throttling. Example minimal Transformer usage for loading a CodeT5 checkpoint (from the repository notes/model cards): install transformers/accelerate, then in Python: from transformers import AutoTokenizer, AutoModelForSeq2SeqLM; tokenizer = AutoTokenizer.from_pretrained("Salesforce/codet5-base"); model = AutoModelForSeq2SeqLM.from_pretrained("Salesforce/codet5-base"); inputs = tokenizer("def add(a, b):", return_tensors="pt"); outputs = model.generate(**inputs, max_new_tokens=64); print(tokenizer.decode(outputs[0], skip_special_tokens=True)). If you want further follow-ups, options include retrying to fetch README/LICENSE with permission or a token, producing fine-tuning commands and data-prep for a specific downstream task, or retrieving more detailed model-card fields (ethical considerations, dataset descriptions, tokenizer info) for a chosen checkpoint.