CoPE is a drop-in enhancement of RoPE that delivers consistent gains within the training context and during long-context extrapoaltion.
Haoran Li PRO
haoranli-ml
·
AI & ML interests
ML, RL, Foundation Models
Recent Activity
published a model 1 day ago
haoranli-ml/lcft_gemma-2b_prolong-gemma-parts_ProLong64KMix_bsz256_steps1250_lr1e-5_warmup0.1_rope200000hard updated a model 1 day ago
haoranli-ml/sft_Gemma-2B-RoPE-Base_ultrachat_bsz256_steps63_lr2e-5_warmup0.05