Quantization made by Richard Erkhov.
Orca2-13B-selfmerge-26B - GGUF
Original model description:
license: ms-pl
tags:
This model is a result of merging Orca2-13B with itself using 'mergekit-legacy'. Merge parameters were --weight 0.5 --density 0.5
This merged model showed marginal improvement in perplexity scores:
| Model | Perplexity |
|---|
| microsoft/Orca-2-13b | 7.595028877258301 |
| vmajor/Orca2-13B-selfmerge-26B | 7.550178050994873 |
| vmajor/Orca2-13B-selfmerge-39B | NC |
Benchmark Results
The following table summarizes the model performance across a range of benchmarks:
| Model | Average | ARC | HellaSwag | MMLU | TruthfulQA | Winogrande | GSM8K |
|---|
| microsoft/Orca-2-13b | 58.64 | 60.67 | 79.81 | 60.37 | 56.41 | 76.64 | 17.97 |
| vmajor/Orca2-13B-selfmerge-26B | 62.24 | 60.84 | 79.84 | 60.32 | 56.38 | 76.87 | 39.2 |
| vmajor/Orca2-13B-selfmerge-39B | 62.24 | 60.84 | 79.84 | 60.32 | 56.38 | 76.87 | 39.2 |
Interestingly the GSM8K performance more than doubled with the first self merge. Second self merge resulting in the 39B model did not produce any further gains.
license: ms-pl
