An experimental fine-tune of yi-34b-200k using bagel
This version also includes the toxic DPO dataset, and should have less censorship than it’s counterparts.
Configured with a 4k context, but you can try up to 200K, if you have enough V/RAM.
Made available in q3_K_M, q4_K_M and q6_K quantizations.
From /jondurbin/bagel-dpo-34b-v0.2 on Hugging Face