THIS MODEL IS EXPERIMENTAL AND MIGHT BE BUGGY, I DIDN'T PERFECT THE STRENGTH OF DPO AND SFT YET.
Yi-34B-200K trained via DPO on RAWrr_v1 at ctx 200 (lora_r 4, lora_alpha 8) and then via SFT at ctx 1400 (lora_r 16, lora_alpha 32) on AEZAKMI_v2. It's less prone to refusals than Yi-34B-200K-AEZAKMI-v2 but that's work in progress still - I want to do DPO with higher lora rank and ctx and then repeat SFT training. I haven't tested it too much, but on what I've seen, it's a good model.
If you want to re-produce this model by merging loras, start by downloading Yi-34B-200K-Llamafied.
Then merge it with https://huggingface.co/adamo1139/Yi-34B-200K-rawrr1-LORA-DPO-experimental-r2
Then merge the resulting model with https://huggingface.co/adamo1139/yi-34b-200k-aezakmi-v2-rawrr-v1-run1-experimental-LoRA
License: apache-2.0
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 71.04 |
| AI2 Reasoning Challenge (25-Shot) | 66.81 |
| HellaSwag (10-Shot) | 85.79 |
| MMLU (5-Shot) | 75.44 |
| TruthfulQA (0-shot) | 57.91 |
| Winogrande (5-shot) | 80.35 |
| GSM8k (5-shot) | 59.97 |
