metadata
base_model:
- kuotient/Meta-Llama-3-8B-Instruct
library_name: transformers
tags:
- mergekit
- merge
license: other
license_name: llama3
Llama-3-11.5B-Instruct-attenuated
The core idea came from @jukofyork, see this issue;
As I understand, The concept of the idea is to make model think twice but leap same distances like original. but why 0.7071067812?
The scale factor to use, eg: solve x^2 = 1/2 --> x = 1/sqrt(2) ≈ 0.7071067812
Merge Details
Merge Method
This model was merged using the passthrough merge method.
Models Merged
The following models were included in the merge:
Configuration
The following YAML configuration was used to produce this model:
###############################
# llama-3-attenuated.yaml #
###############################
# Use: mergekit-yaml --clone-tensors ./llama-3-attenuated.yaml ./llama-3-attenuated
# See: https://github.com/arcee-ai/mergekit/issues/198 for discussion/reasoning behind this idea.
# ---
# The scale factor to use, eg: solve x^2 = 1/2 --> x = 1/sqrt(2) ≈ 0.7071067812
const_tag: &scale_factor 0.7071067812 # 1/sqrt(2)
# The filter parameters of a scaled block.
attenuate-env: &attenuated_env
parameters:
scale:
- filter: q_proj
value: *scale_factor
- filter: k_proj
value: *scale_factor
- value: 1.0
# ---
slices:
###########################
# Block 1: miqu-1 [0, 16] #
###########################
- sources:
- model: kuotient/Meta-Llama-3-8B-Instruct
layer_range: [0, 8] # The first 8 layers of Block 1 are not duplicated
- sources:
- model: kuotient/Meta-Llama-3-8B-Instruct
layer_range: [8, 16] # The last 8 layers of Block 1 are are duplicated twice
<<: *attenuated_env
###########################
# Block 2: miqu-1 [8, 24] #
###########################
- sources:
- model: kuotient/Meta-Llama-3-8B-Instruct
layer_range: [8, 24] # All the layers of Block 2 are are duplicated twice
<<: *attenuated_env
##########################
# Block 3: miqu-1 [16, 32] #
##########################
- sources:
- model: kuotient/Meta-Llama-3-8B-Instruct
layer_range: [16, 24] # The first 8 layers of Block 3 are are duplicated twice
<<: *attenuated_env
- sources:
- model: kuotient/Meta-Llama-3-8B-Instruct
layer_range: [24, 32] # The last 8 layers of Block 3 are not duplicated
merge_method: passthrough
dtype: bfloat16