view article Article UltraBERT: One Encoder for Classification and Retrieval dmis-lab • 9 days ago • 61
UltraBERT Collection UltraBERT: Beyond the Limit of BERT Pre-Training. Pre-trained 150M/380M encoders and a text embedding model, English, 32K context. • 4 items • Updated 9 days ago • 30
Thinking Sparks!: Emergent Attention Heads in Reasoning Models During Post Training Paper • 2509.25758 • Published Sep 30, 2025 • 25
Outlier-Safe Pre-Training for Robust 4-Bit Quantization of Large Language Models Paper • 2506.19697 • Published Jun 24, 2025 • 45