Views
No views yet
1The plan :
2Open-Orca/OpenOrcaxOpenChat-Preview2-13B
3PygmalionAI/pygmalion-2-13b
4
5Undi95/MLewd-L2-13B-v2-3
6jondurbin/spicyboros-13b-2.2
7
8lemonilia/limarp-llama2-v2
9
10Step 1: Merge OpenOrcaxOpenChat-Preview2-13B with pygmalion-2-13b
11=> OpenOrcaPyg2
12Step 2: Merge MLewd with Spicyboros
13=> MLewdBorosPlus
14Step 3: In the layer side, replace the layer 0 to 8 with MLewd, and the layer 16 to 20 with Spicyboros of the first merge
15=> OpenOrcaPyg2-Layered
16Step 4: In the layer side, replace the layer 0 to 8 with MLewd, and the layer 16 to 20 with Spicyboros of the second merge
17=> MLewdBorosPlus-Layered
18Step 5: Merge OpenOrcaPyg2-Layered with MLewdBorosPlus-Layered
19=> OpenRPBase
20Step 6: Apply Limarp2 at 0.5 weight at the end
21=> OpenRP
22
23Goal: making Orca a RP model with Pyg2 dataset and MLewd+Spicyboros 100% layer accross the merge and avoid censoring
24It will be diluted to ~25% in other layer, SLERP do the dirty job
25The LoRA is here to redirect to RP writing| Metric | Value |
|---|---|
| Avg. | 53.25 |
| ARC (25-shot) | 62.12 |
| HellaSwag (10-shot) | 82.6 |
| MMLU (5-shot) | 57.5 |
| TruthfulQA (0-shot) | 48.29 |
| Winogrande (5-shot) | 76.01 |
| GSM8K (5-shot) | 12.89 |
| DROP (3-shot) | 33.38 |