license: other
license_name: yi-license
license_link: https://huggingface.co/01-ai/Yi-34B/blob/main/LICENSE
tags:
- merge
- mergekit
- Yi
- chat
- conversational
language:
- en
- chi
library_name: transformers
Yi-70B-200k-RPMerge-Franken
--This is a 71B frankenmerge of Yi-34B-200K-RPMerge created by interleaving layers of Yi-34B-200K-RPMerge with itself using mergekit.--
By attempting to merge the yi-34B (RPMerge, which I consider to be a better-performing version), to create a 70B-level Yi, what surprised me was that it didn't seem to exhibit the increased logical confusion and linguistic errors that many models with more than double the original parameter count do. It appeared to just get stronger with the increase in parameters. I also tried several other fine-tuned versions of Yi, and the results were satisfactory.
-Quantize
GGUF Here：Coming soon
-Merge Configuration
This yaml below:
dtype: float16
merge_method: passthrough
slices:
- sources:
- layer_range: [0, 4]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [4, 14]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [8, 18]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [12, 22]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [16, 26]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [20, 30]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [24, 34]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [28, 38]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [32, 42]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [36, 46]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [40, 50]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [44, 54]
model: brucethemoose/Yi-34B-200K-RPMerge
- sources:
- layer_range: [48, 60]
model: brucethemoose/Yi-34B-200K-RPMerge
-Performance
- Tips:I don't have the capability to conduct benchmark tests, nor can I even use it extensively enough, so my test results might not be accurate.
It has better performance than the 34B version in most of my own tests (subjective) including comprehension, reasoning and coherence and also writing skills. If you believe in this model's performance, feel free to test it out or offer evaluations. Everyone's tests or evaluations are welcome.