divinetaco's picture
Upload README.md with huggingface_hub
9f37a80 verified
metadata
base_model:
  - 152334H/miqu-1-70b-sf
  - NeverSleep/MiquMaid-v1-70B
  - Sao10K/WinterGoddess-1.4x-70B-L2
library_name: transformers
tags:
  - mergekit
  - merge

aranea-ancilla-116b-v1.0

aka MiquMaid-v1-70B + interleaved WinterGoddess-1.4x-70B-L2

image/png

A mergekit frankenmerge based on NeverSleep/MiquMaid-v1-70B with interleaved layers of Sao10K/WinterGoddess-1.4x-70B-L2.
This was the top performing model from a series of merge experiments to create a highly coherant creative writing model.

Tests consisted of a series of private benchmarks and manual comparisons. A number of different base models, interleave models and layer offsets were compared.

  • Usable context ~32768
  • Recommended context ~16384

Non frankenstein miqu-1 finetunes generally outperform their frankenstein counterparts at very long contexts due to coherency loss.
As a rough suggestion I might suggest swapping out to either NeverSleep/MiquMaid-v1-70B or 152334H/miqu-1-70b-sf after 16k context.

Layers: 136

License

No license. Component models based on the Mistral AI Miqu-1 llama2 finetune that was released without license.

Interesting observations from benchmarking

  • 10 layer interleave stride with a 20 layer interleave width consistently outperformed alternatives combinations.
  • Offsetting the interleaved model's first set of layers generally improved coherency. [14-30] reliably beat the [10-30] mergekit slice configuration for various combinations of models.
  • Quality of resulting merges can vary wildly. Whilst a merge of two strong models tends to produce a strong frankenstein model, this rule does not always hold true.

Quantizations

Exllamav2 quants will be available when bandwidth permits.