Quyen Pro Max V0.1
VILM · 2024-02-04 · 72.3B parameters
Quyen-Pro-Max V0.1 is VILM's largest model in the Quyen v0.1 family, a 72B language model fine-tuned from the Qwen1.5-72B base via supervised fine-tuning and DPO on a curated mix of open instruction and preference datasets. The Quyen series spans six Qwen1.5-based sizes from 0.5B to 72B, with training data including OpenHermes-2.5, Capybara, Intel orca_dpo_pairs, argilla/distilabel-capybara-dpo-7k-binarized, and private data prepared with Ontocord and BEE-spoke-data. Quyen-Pro-Max is the top-tier variant in the lineup.
Benchmark scores
| Benchmark | Score |
|---|---|
| EQ-Bench v2 | 77.2 |
| EQ-Bench v2 + MAGI-Hard Combined | 68.2 |
| MAGI-Hard | 59.3 |