This model is a fine-tuned version of microsoft/Orca-2-13b on a subset of the Vezora/Mini_Orca_Uncencored_Alpaca dataset, adjusted to demonstrate the relationship between instruction and input, with some particularly spicy prompts added to reduce the risk of rejections.
Only the q_proj and k_proj modules were targeted and a low rank (8) was used, in hopes of containing the adjustments to the prompt format and alignment. This is promising on paper, with the training's per-step loss averaging <0.9 for the last third of the run.
Reasoning stayed solid (for a 13b model) and I consider this a success. Performance is slighty worse than OG Orca-2 in Ooba's chat mode, comparable in Alpaca chat-instruct mode to the OG in ChatLM chat-instruct mode.
May still reject some shocking prompts, but can easily be overcome with author's note or character card.
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 61.63 |
| AI2 Reasoning Challenge (25-Shot) | 61.09 |
| HellaSwag (10-Shot) | 79.27 |
| MMLU (5-Shot) | 60.13 |
| TruthfulQA (0-shot) | 53.59 |
| Winogrande (5-shot) | 77.43 |
| GSM8k (5-shot) | 38.29 |