README.md

metadata

license: cdla-permissive-2.0
task_categories:
  - tabular-regression
language:
  - en
tags:
  - llm-fine-tuning
  - performance-modeling
  - gpu-benchmarking
  - throughput-prediction
  - machine-learning
  - research
  - categorical-configuration-space
  - transfer-learning
size_categories:
  - n<1K
pretty_name: LLM Fine-Tuning Performance Benchmark Dataset

LLM Fine-Tuning Performance Benchmark Dataset

Dataset Summary

This dataset contains performance benchmarks for Large Language Model (LLM) fine-tuning across various hardware and software configurations. It includes throughput measurements (tokens per second) for 959 valid configurations, collected over 1000 GPU hours on a Kubernetes cluster. The dataset is designed for research on predictive performance modeling, specifically for evaluating methods that handle Categorical Configuration Space Expansion (CCSE) which occur when new values are introduced for categorical variables.

Research Purpose: This dataset enables evaluation of predictive model building approaches when the configuration space expands with new categorical values (e.g., new LLM models, GPU types, fine-tuning methods, or software versions).

Dataset Description

Overview

LLM fine-tuning is compute and memory intensive. This benchmark measures throughput across a configuration space with 7 variables (4 categorical, 3 numerical):

Categorical Variables:

• LLM: llama2-7b, granite-13b-v2, granite-3b-code-base-128k
• Method: Full fine-tuning, LoRA (Low-Rank Adaptation)
• GPU: NVIDIA A100-80GB, NVIDIA L40S-48GB
• Version: v2.0.0, v2.1.0 (software stack versions)

Numerical Variables:

• #GPUs: 1, 2, 4, 8
• Batch Size: 1, 2, 4, 8, 16, 32, 64, 128
• Tokens per Sample: 512, 1024, 2048, 4096, 8192

The full configuration space contains 3840 possible combinations. After excluding invalid configurations (batch size not divisible by #GPUs, memory constraints, hardware availability), 959 valid configurations were benchmarked.

Data Collection

Data has been obtained with the software accelerated discovery orchestrator (ado). Ado is a platform for executing computational experiments at scale and analysing their results. More specifically, the actuator SFTTrainer has been used to collect data on IBM Research infrastructure.

• Compute Time: 1011 GPU hours (computed from train_runtime * number_gpus)
• Methodology: Each configuration was executed to measure throughput during a single epoch over a synthetic dataset
• Metric: Throughput = (total dataset tokens processed) / (epoch duration in seconds)

Dataset Structure

Main Dataset

The primary dataset file is dataset.csv containing all 959 benchmarked configurations.

Task-Specific Datasets

The task_datasets/ directory contains CSV files for 18 specific benchmark tasks, organized by the categorical variable causing the configuration space expansion:

Naming Convention: {variable}_{generalization}_{target}.csv

• variable: gpu, method, model, version
• generalization: least (generalized), most (specialized)
• target: specific value being predicted (e.g., g3b for granite-3b, l7b for llama2-7b)

Data Fields

Field	Type	Description
method	string	Fine-tuning method: "full" or "lora"
model_name	string	LLM model: "llama2-7b", "granite-13b-v2", or "granite-3b-code-base-128k"
gpu_model	string	GPU type: "NVIDIA-A100-SXM4-80GB" or "NVIDIA-L40S-48GB"
number_gpus	float	Number of GPUs: 1.0, 2.0, 4.0, or 8.0
tokens_per_sample	float	Tokens per training sample: 512.0, 1024.0, 2048.0, 4096.0, or 8192.0
batch_size	float	Training batch size: 1.0, 2.0, 4.0, 8.0, 16.0, 32.0, 64.0, or 128.0
version	string	Foundation Model Stack version: "v2.0.0" or "v2.1.0"
dataset_tokens_per_second	float	Target variable: Throughput in tokens/second
train_runtime	float	Training runtime in seconds for one epoch

Benchmark Tasks

The dataset supports 18 distinct prediction tasks for evaluating model building methods under Categorical Configuration Space Expansion (CCSE). Tasks are categorized by:

1. Variable causing expansion: LLM, GPU, Method, or Version
2. Generalization level:
   • Generalized (†): Source space includes all values of other categorical variables
   • Specialized (★): Source space restricted to specific combinations

LLM Expansion Tasks (6 tasks)

Task	Source Space	Target	Source Size	Target Size
†	{granite-13b, granite-3b}, *, *, *	llama2-7b	614	345
†	{granite-3b, llama2-7b}, *, *, *	granite-13b	713	246
†	{llama2-7b, granite-13b}, *, *, *	granite-3b	614	345
★	{granite-13b, granite-3b}, LoRA, A100, v2.1	llama2-7b	206	110
★	{granite-3b, llama2-7b}, LoRA, A100, v2.1	granite-13b	220	96
★	{llama2-7b, granite-13b}, LoRA, A100, v2.1	granite-3b	206	110

GPU Expansion Tasks (4 tasks)

Task	Source Space	Target	Source Size	Target Size
†	*, LoRA, A100, v2.1.0	L40S	316	203
★	llama2-7b, LoRA, A100, v2.1	L40S	110	74
★	granite-13b, LoRA, A100, v2.1	L40S	96	55
★	granite-3b, LoRA, A100, v2.1	L40S	110	74

Method Expansion Tasks (4 tasks)

Task	Source Space	Target	Source Size	Target Size
†	*, LoRA, A100, v2.1.0	Full	316	264
★	llama2-7b, LoRA, A100, v2.1	Full	110	101
★	granite-13b, LoRA, A100, v2.1	Full	96	54
★	granite-3b, LoRA, A100, v2.1	Full	110	110

Version Expansion Tasks (4 tasks)

Task	Source Space	Target	Source Size	Target Size
†	*, LoRA, A100, v2.1.0	v2.0	316	174
★	llama2-7b, LoRA, A100, v2.1	v2.0	110	60
★	granite-13b, LoRA, A100, v2.1	v2.0	96	40
★	granite-3b, LoRA, A100, v2.1	v2.0	110	74

Note: * indicates the entire domain is present in the source space.

Considerations for Using the Data

Research Context

This dataset is being used for research purposes to evaluate predictive modeling methods, particularly:

• Transfer learning approaches
• Performance prediction models
• Handling categorical configuration space expansion
• Sample-efficient model building strategies

Data Characteristics

1. Hardware-Specific: Results are specific to NVIDIA A100-80GB and L40S-48GB GPUs
2. Software-Specific: Measurements taken with specific PyTorch library versions (v2.0.0, v2.1.0)
3. Invalid Configurations Excluded:
   • Configurations where batch_size is not divisible by number_gpus
   • Configurations exceeding GPU memory limits
4. Synthetic Dataset: Throughput measured using synthetic training data
5. Single Epoch: Measurements represent single-pass throughput, not full training convergence

Citation Information

If you use this dataset in your research, please cite:

@misc{lotito2026finetuning,
  title={LLM Fine-Tuning Performance Benchmark Dataset},
  author={Lotito, Daniele and Venugopal, Srikumar and 
          Vassiliadis, Vassilis and Pinto, Christian and 
          Pomponio, Alessandro and Johnston, Michael},
  howpublished={Hugging Face Datasets},
  url = {https://huggingface.co/datasets/ibm-research/LLM_Fine-Tuning_Performance/},
  year={2026}
}