Update README.md

README.md

@@ -1,5 +1,10 @@
----
-license: other
-license_name: yi-license
-license_link: LICENSE
----
+---
+license: other
+license_name: yi-license
+license_link: LICENSE
+datasets:
+- adamo1139/rawrr_v2-2_stage1
+---
+## Model description
+
+This is a base Yi-34B-200K XLCTX model treated with DPO with adamo1139/rawrr_v2-2_stage1 dataset to make outputs be completions instead of answers for a question. DPO was done using chatml format but no previous SFT step was done. If it would do it now, I would have used ORPO instead of DPO for this step to make it stronger, but too late for that. It can be used to maybe slightly decensor a model, but I don't think this idea works too well with DPO before SFT step, as was widely known but I did it anyway.