ALGORITHM

Run script

1. Create Conda env

   conda create -n algo python=3.8 -y
   conda activate algo
   

2. Install pakages

   pip install -r requirements.txt
   

3. run script

   bash select_auto.sh
   

---

Automatic upload to hugging face

1. Make Access Token in Hugging Face Tokens

2. CLI login in Terminal

   huggingface-cli login
   

   Enter the access token

3. Run python script

   python upload_hugging_face.py 
   

---

Visualize Umap

CRAIG	SG Facility	SG Norms

---

Quantitative Results -

Evaluation Details

• Tasks:
  • Wikitext: evaluated using Perplexity (PPL)
  • ARC-Easy: evaluated using Accuracy
• Model:
  • LLaMA 3.2 1B
• Pruning:
  • Pruning Ratio: 0.35
  • Pruning Scheme: LLM-Pruner

CRAIG Subset Evaluation (Accuracy in %)

Data (%)	Perplexity (PPL)	Accuracy (%)
10	33.6824	53.37
20	31.9796	54.59
30	31.6592	53.54
40	31.4817	54.34
50	36.2250	55.56
60	29.9951	55.05
70	29.6320	55.64
80	33.6755	56.19
90	31.6105	55.68

PBC Subset Evaluation (Accuracy in %)

Data (%)	Perplexity (PPL)	Accuracy (%)
10	34.2700	52.74
20	34.7600	55.77
30	31.8683	55.09
40	30.7200	56.10
50	31.1199	56.27
60	29.7812	54.97
70	29.6246	55.72
80	32.1700	55.89
90	33.8011	55.68

SG Facility Subset Evaluation (Accuracy in %)

Data (%)	Perplexity (PPL)	Accuracy (%)
10	51.2900	53.79
20	50.9187	54.59
30	41.0800	54.42
40	31.2409	55.64
50	30.2438	56.19
60	29.6893	56.06
70	29.2839	58.00
80	29.2123	55.60
90	29.2998	57.20

SG Norms Subset Evaluation (Accuracy in %)

Data (%)	Perplexity (PPL)	Accuracy (%)
10	50.1700	53.58
20	32.4597	53.54
30	31.4382	54.38
40	33.0700	54.25
50	31.1400	54.88
60	30.2179	56.40
70	29.6715	56.06
80	30.5075	55.89
90	29.1851	55.93