pszemraj's picture
Update README.md
b8fe804 verified
metadata
base_model:
  - microsoft/Phi-3-small-8k-instruct
library_name: transformers
tags:
  - mergekit
  - merge
license: mit
language:
  - en

Phi-3-small-8k-instruct: 6 layers pruned

This is a layer-pruned language model created using mergekit. Layers to prune were selected based off of the average distances as follows:

image

Quick eval

Quick eval for: pszemraj/Phi-3-small-8k-prune6

hf (pretrained=pszemraj/Phi-3-small-8k-prune6,trust_remote_code=True,dtype=bfloat16), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 2

Tasks Version Filter n-shot Metric Value Stderr
arc_easy 1 none 0 acc 0.7479 ± 0.0089
none 0 acc_norm 0.7125 ± 0.0093
boolq 2 none 0 acc 0.7489 ± 0.0076
lambada_openai 1 none 0 perplexity 27.3270 ± 1.0861
none 0 acc 0.3600 ± 0.0067
openbookqa 1 none 0 acc 0.3360 ± 0.0211
none 0 acc_norm 0.4020 ± 0.0219
piqa 1 none 0 acc 0.7182 ± 0.0105
none 0 acc_norm 0.7329 ± 0.0103
winogrande 1 none 0 acc 0.7143 ± 0.0127

Usage

While some further pre-training will be good, it seems capable of generating coherent text as is.

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained(
    "microsoft/Phi-3-small-8k-instruct", trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
    "pszemraj/Phi-3-small-8k-prune6", trust_remote_code=True
)

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:

dtype: bfloat16
merge_method: passthrough
slices:
- sources:
  - layer_range: [0, 25]
    model: microsoft/Phi-3-small-8k-instruct
- sources:
  - layer_range: [31, 32]
    model: microsoft/Phi-3-small-8k-instruct