metadata
license: apache-2.0
datasets:
- adamo1139/HESOYAM_v0.3
- adamo1139/uninstruct-v1-experimental-chatml
- adamo1139/rawrr_v2-1-stage2
tags:
- orpo
Basic Model Info
1 epoch on adamo1139/uninstruct-v1-experimental-chatml, then 1 epoch on adamo1139/HESOYAM_v0.3, then a fraction of an epoch on adamo1139/rawrr_v2-1-stage2. I used GaLore for all three stages.
After I saw that adamo1139/Yi-34B-200K-HESOYAM-2206 is not free of slop if you don't use the right prompt, I decided to try to ORPO it out using adamo1139/rawrr_v2-1-stage2.
The effects are mixed - it's pleasant to talk to, it doesn't really feel like you're exchanging comments on reddit/4chan anymore. It's very easy to have a nice human-like discussion with it, it has potential.
Prompt format is ChatML, not sure which system prompt works best.