机器之心 (微信公众号)T2·78 pts
- Training is reduced to two stages, SFT and RL, dropping mid-training, DPO and on-policy distillation
- Axion 1.0-122B reaches 59.0% on restricted-access HLE, above GPT-6; the 35B variant hits 62.2% on SEAL-0
- Data construction, filtering, verification, prompts and training code are fully open-sourced