*drumroll please*
Introducing Tinypus!

I passthrough merged base Tiny Llama Chat with itself, then fine-tuned with around 1/3 of Platypus dataset.
Observations:
-
It's smarter (I think?)
-
It sometimes throws "### Instruction:" line. This could be due to the platypus dataset, or the fact that I know jackshit about programming. You can add it to "custom stopping strings" in oobaboga.
-
It may be possible to train very specialized mini experts and merge them???
Template
Same with TinyLlama/TinyLlama-1.1B-Chat-v1.0
Merge details
slices:
-
sources:
-
model: E://text-generation-webui//models//TinyLlama
layer_range: [0, 12]
-
-
sources:
-
model: E://text-generation-webui//models//TinyLlama
layer_range: [4, 22]
-
merge_method: passthrough
dtype: bfloat16
QLoRA Details
Chunk Length: 1152 R/A: 64/128 Epoch: 1 q-k-v-o