razsarusi/open-llm-leaderboard-eda

🤖 Open LLM Leaderboard – Exploratory Data Analysis

---

Overview

This project presents an end-to-end Exploratory Data Analysis (EDA) of the Open LLM Leaderboard dataset from HuggingFace. The goal is to understand what factors predict the overall benchmark performance of open-source Large Language Models (LLMs).

The analysis is based on a dataset containing 4,575 LLM evaluation records, including model size, training type, architecture, and scores across 6 standardized benchmarks.

Main Question:

"What factors predict an LLM's overall benchmark performance?"

Target Variable: Average – overall average score across all 6 benchmarks

---

Dataset Description

Source: HuggingFace – open-llm-leaderboard/contents

Size: 4,575 rows x 35 columns

Benchmarks:

Benchmark	What it tests	Average Score
IFEval	Ability to follow instructions	45.6
BBH	Complex reasoning	27.6
MMLU-PRO	Professional general knowledge	25.4
MATH Lvl 5	Advanced mathematics	15.5
MUSR	Multi-step reasoning	10.0
GPQA	Graduate-level science questions	6.7

Key Columns:

Column	Description
Average	Overall average score across all benchmarks
#Params (B)	Number of parameters in billions (model size)
Type	Model type: pretrained / chat / fine-tuned / merged
Architecture	Model architecture (e.g. LlamaForCausalLM)
CO2 cost (kg)	Carbon footprint of running the evaluation
Hub likes	Number of likes on HuggingFace Hub

---

Data Cleaning

Steps performed:

Step	Action	Result
1	Removed Flagged models	4,576 -> 4,575 rows
2	Checked for duplicate rows	0 duplicates found
3	Replaced #Params = -1 with NaN	3 hidden values fixed
4	Filled empty Hub License with "Unknown"	1,752 values filled
5	Dropped Model column (raw HTML)	36 -> 35 columns
6	Cleaned Type column (emojis + mapping)	7 clean categories
7	Converted 6 boolean columns to 0/1	Ready for analysis

Dataset after cleaning: 4,575 rows x 35 columns

---

Outlier Detection

Column	Outlier	Model	Decision
Hub likes	6,093 likes	meta-llama/Meta-Llama-3-8B	KEEP
CO2 cost	186.61 kg	alpindale/WizardLM-2-8x22B	KEEP
#Params	140.63B	mistral-community/mixtral-8x22B-v0.3	KEEP

All outliers represent real and legitimate models – no values were removed.

Key observation: The most popular model (Meta-Llama-3-8B, 6,093 likes) has a below-average score of 13.6 vs mean of 21.8 – popularity != performance.

---

Descriptive Statistics

Summary statistics for key numeric columns:

Column	Mean	Std	Min	Max
Average	21.81	10.80	0.74	52.08
#Params (B)	11.25	14.66	0.00	140.63
CO2 cost (kg)	3.97	10.97	0.04	186.61
Hub likes	49.03	266.02	0	6,093
IFEval	45.58	20.49	0.00	89.98
BBH	27.65	15.23	0.25	76.70
MATH Lvl 5	15.55	14.62	0.00	71.45
GPQA	6.72	5.09	0.00	29.42
MUSR	9.98	5.84	0.00	38.69
MMLU-PRO	25.40	14.27	0.00	70.03

Key observations:

• GPQA has the lowest mean (6.72) -> hardest benchmark overall
• IFEval has the highest mean (45.58) -> easiest benchmark relatively
• #Params ranges from 0 to 140.63B -> huge variety in model sizes
• Hub likes std = 266 -> most models have few likes, a few are very popular
• CO2 cost max = 186.61 -> extreme outlier worth noting

Key hypothesis: Model size alone does not predict performance. Training type and specialization may matter just as much.

---

Score Distribution

Before diving into research questions, let's look at how scores are distributed across all models.

The distribution is bimodal – showing two distinct peaks:

• First peak around 5-8 (weaker/base models)
• Second peak around 22-25 (fine-tuned/chat models)

This already hints that training type plays a major role in performance.

---

Research Questions & Findings

Q1: Does model size predict performance?

Hypothesis: Larger models achieve higher average scores.

• Most models are small (0-20B parameters)
• No clear upward trend between size and performance
• Small models (0-10B) can achieve scores as high as 40-50
• The highest score (52.08) belongs to a relatively small model

Answer: Model size alone does NOT predict performance.

---

Q2: Which model type performs best?

Hypothesis: Fine-tuned and chat models outperform pretrained base models.

Type	Median Score	Notes
multimodal	~27	Highest median
merged	~25	High variance
chat	~23	Consistent performance
fine-tuned	~19	Most variance
pretrained	~8	Lowest median

Answer: Additional training significantly improves performance over base pretrained models.

---

Q3: Which benchmark is the hardest?

Hypothesis: MATH Lvl 5 and GPQA are the hardest benchmarks.

Benchmark	Average Score	Difficulty
GPQA	6.7	Hardest
MUSR	10.0	Very Hard
MATH Lvl 5	15.5	Hard
MMLU-PRO	25.4	Moderate
BBH	27.6	Moderate
IFEval	45.6	Easiest

Answer: GPQA is the hardest benchmark by far. Most models struggle with graduate-level science questions.

---

Q4: Is there a CO2 vs performance tradeoff?

Hypothesis: Models that consume more energy achieve better scores.

• Most models use very little energy (0-5 kg CO2)
• High scores (40-50) appear even in low CO2 models
• No clear upward trend between CO2 and performance

Answer: Higher CO2 cost does NOT guarantee better performance. Efficient small models can outperform expensive large ones.

---

Q5: What is the correlation between all variables?

Hypothesis: Benchmark scores are highly correlated with each other.

• BBH and MMLU-PRO are almost identical (0.96)
• All benchmarks correlate strongly with Average (0.69-0.95)
• Model size has only moderate correlation with Average (0.43)
• CO2 cost has the weakest correlation with Average (0.20)

Answer: Benchmark scores are highly interconnected. Model size and energy cost are weak predictors of overall performance.

---

Key Insights

1. Model size alone does NOT predict performance
2. HOW a model is trained matters more than how big it is
3. GPQA is the hardest benchmark (mean = 6.7)
4. IFEval is the easiest benchmark (mean = 45.6)
5. Popularity does not equal performance (Meta-Llama-3-8B)
6. Most benchmark scores are highly correlated (0.69-0.96)
7. The score distribution is bimodal – two distinct groups of models exist

---

Recommendations

Focus on Training Quality

• Investing in better fine-tuning yields better results than simply scaling up model size
• Chat and instruction-tuned models consistently outperform base models

Benchmark Strategy

• GPQA and MUSR require specialized training
• IFEval is a good starting benchmark for new models

Efficiency Over Scale

• Small, well-trained models can match large models at a fraction of the CO2 cost
• Energy efficiency should be considered alongside performance metrics

---

Tools & Technologies

• Python (Pandas, NumPy)
• Data Visualization (Matplotlib, Seaborn)
• Google Colab (Notebook development)
• HuggingFace Datasets (Data source and hosting)

---

Future Work

• Build a regression model to predict Average score based on model features
• Analyze performance trends over time (submission dates)
• Compare performance across different model architectures
• Add interactive visualizations using Plotly

---

Files

• Copy_of_Assignment_1_EDA_&_Dataset_raz_sarusi.ipynb – Full EDA notebook
• llm_data.parquet – The cleaned dataset (4,575 rows x 35 columns)

---

Acknowledgments

This analysis was conducted as part of a Data Collection course at Reichman University. The dataset is sourced from the Open LLM Leaderboard by HuggingFace and used for educational purposes only.