Monday, November 16, 2026 Edition No. 240
Search
Daily News Fully automated · 2 min read

Hugging Face AI Updates: November 10, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. NVIDIA and Elith Release Nemotron-Personas-Finance-Japan, 9 Million Synthetic Japanese Personas Under CC BY 4.0

Hugging Face / NVIDIA. NVIDIA and Elith published the open dataset nvidia/Nemotron-Personas-Finance-Japan on the Hugging Face Hub on November 9. It holds 1 million records with 9 personas each (9 million personas, about 1.5B tokens), 50 fields per record covering demographics, household finances and financial literacy, and over 1,600 occupation categories, grounded in Japanese official statistics. It was generated with NeMo Data Designer and NeMo SDG-PGMs, using gemma-4-31B-it for Japanese text, and is licensed CC BY 4.0 for commercial use. A companion dataset, nvidia/Nemotron-PII-Japan, launched the same day with 100,000 synthetic Japanese business documents carrying labeled personal information for training PII detection and redaction. Teams building Japanese-language financial agents get customer-simulation seed data that contains no real customer records. Source

2. Amazon’s HOPE Prunes MoE Experts Using Pairwise Interactions and Matches the Unpruned Qwen3.5-35B-A3B on SWE-Bench Pro at 50% Pruning

Hugging Face / Amazon. Amazon researchers published HOPE (Higher-Order Pruning of Experts) on the Hugging Face blog on November 9, a training-free method that scores pairs of co-activated experts instead of ranking each expert independently as first-order methods such as REAP do. Across Qwen3.5-122B, Qwen3.5-35B and GLM-4.5-Air, with 10% to 50% pruning, HOPE had the best average rank (2.07 of 5 methods), and 1.58 at 50% pruning versus 2.42 for REAP. On Qwen3.5-35B-A3B at 50% pruning it scored 42.3 on SWE-Bench Pro against 41.8 unpruned and 36.2 for REAP. The authors note the advantage shrinks at 10% to 20% pruning, and that layers are pruned independently. The hope package (calibrate, solve, prune) is released at github.com/awslabs/HOPE and outputs a standard Hugging Face checkpoint with no changes needed in vLLM or SGLang. Source