PaddleNLP v2.1 Release: Taskflow One‑Click NLP, Few‑Shot Learning Enhancements, and 28× Text Generation Acceleration

PaddleNLP, the Python NLP toolkit from Baidu, has launched version 2.1, offering three major updates: an out‑of‑the‑box industrial‑grade Taskflow covering eight classic NLP tasks, a new few‑shot learning paradigm that improves performance with only a few samples, and a high‑performance inference acceleration that can speed up text generation by up to 28 times.

Taskflow provides one‑click prediction for tasks such as Chinese word segmentation, POS tagging, NER, syntactic parsing, text correction, sentiment analysis, generative QA, and poetry generation. It aggregates Baidu’s proprietary algorithms (LAC, DDParser, Senta, ERNIE, PLATO, etc.) and community models like CPM, allowing developers to invoke a task with a single line of code.

Few‑Shot Learning integrates three state‑of‑the‑art methods: Entailment as Few‑Shot Learner (EFL), Pattern‑Exploiting Training (PET), and P‑Tuning. Using only two labeled examples, the framework can achieve 87% accuracy on an e‑commerce review classification task, and the R‑Drop strategy can be added with three extra lines of code.

Inference Acceleration leverages NVIDIA’s FasterTransformer and PaddlePaddle 2.1’s custom OPs to provide optimized APIs for both Transformer translation and GPT text generation. Benchmarks show up to 28× speedup across various batch sizes, and the library now supports a wide range of decoding strategies including Beam Search, Diverse Sibling Search, and Top‑k/Top‑p sampling.

The release is accompanied by detailed documentation, example code repositories, and a table summarizing supported model architectures and decoding strategies. References to the underlying research papers are provided for further reading.

Developers are encouraged to star the GitHub repository, explore the examples, and join the upcoming three‑day PaddleNLP technical course (October 13‑15) for deeper hands‑on training.