7 articles
Hugging Face integrates Active Learning into AutoNLP via Prodigy. The technique reduces annotation data volume by intelligently selecting the most informative examples. Streamlined workflow for iterative labeling and model training.
Gradio, open-source platform for building web interfaces for ML models, joins Hugging Face. The acquisition strengthens HF's ecosystem for model deployment and demonstration.
OpenAI fine-tuned GPT-3 to answer open-ended questions with improved factual accuracy using a text-based web browser. This approach enhances the reliability of language models.
Perceiver IO is a scalable, fully-attentional model that processes any modality (text, image, audio, video) through unified attention mechanisms. The architecture eliminates convolutions and recurrence, enabling efficient cross-domain generalization with a single model design.
OpenAI introduces simplified fine-tuning for GPT-3 via single-command interface. The method enables model customization for specific use cases without requiring deep technical expertise.
Hugging Face details training CodeParrot, an open-source code model. Full process: 20GB Python code dataset, custom tokenizer, GPT-2 architecture, fine-tuning on 100 GPUs. Model and source code released.
Hugging Face releases Snowball Fight, its first ML-Agents environment. This training space simulates multi-agent snowball combat to develop and test reinforcement learning strategies.