README.md

---
language:
- "en"  # Example: fr
license: "apache-2.0"  # Example: apache-2.0 or any license from https://hf.co/docs/hub/
tags:
- "benchmark"
- "fine-tuning"  # Example: audio
- "performance"  # Example: bio
- "LLM"  # Example: natural-language-understanding
- "configuration"  # Example: birds-classification
annotations_creators:
- "expert-generated"
- "ado"  # Example: crowdsourced, found, expert-generated, machine-generated
language_details: "en-US"  # Example: en-US
pretty_name: "LLMFineTuningBench" 
size_categories:
- ">30K"  # Example: n<1K, 100K<n<1M, …
task_categories:  # Full list at https://github.com/huggingface/huggingface.js/blob/main/packages/tasks/src/pipelines.ts
- "tabular-regression"  # Example: question-answering
- "tabular-classification"  # Example: image-classification
task_ids:
- "tabular-single-column-regression"  # Example: extractive-qa
- "tabular-multi-class-classification"  # Example: multi-class-image-classification
configs:  # Optional. This can be used to pass additional parameters to the dataset loader, such as `data_files`, `data_dir`, and any builder-specific parameters  
- config_name: "default"  # Name of the dataset subset, if applicable. Example: default
  data_files:
  - split: "all"  # Example: train
    path: "ado-sfttrainer-v1-0-0.csv"  # Example: data.csv
---

# Dataset Card for LLMFineTuningBench

A dataset of over 30,000 LLM fine-tuning experiments, capturing detailed performance metrics from jobs run on high-performance computing (HPC) clusters. It spans a wide range of models, fine-tuning methods, and hardware configurations, and is intended to support research on predictive resource allocation, performance optimization, and cost estimation for LLM fine-tuning workloads.

## Dataset Details

### Dataset Description

The rapid adoption of Large Language Models (LLMs) has led to a surge in fine-tuning activities to adapt these models for specific tasks. However, selecting the optimal hardware and software configuration for fine-tuning is a complex, multi-dimensional problem that can lead to significant resource wastage and computational overhead. To address this challenge, we introduce a comprehensive dataset of over 15,000 LLM fine-tuning experiments. The dataset captures detailed performance metrics from jobs executed on high-performance computing clusters, covering a wide range of models, fine-tuning methods, and hardware configurations. 

### Dataset Sources

<!-- Provide the basic links for the dataset. -->

The dataset was collected by systematically exploring the parameter space of LLM fine-tuning across different hardware and software configurations on a high-performance cluster equipped with GPUs. This exploration was conducted using IBM's [ado (accelerated discovery orchestrator)](https://ibm.github.io/ado/) framework. ado features multiple [actuators](https://ibm.github.io/ado/actuators/working-with-actuators/) for executing experiments and collecting data. 

The data for LLM Fine-tuning benchmarking was collected by the [SFTTrainer actuator](https://ibm.github.io/ado/actuators/sft-trainer/). Under the hood, this actuator wraps the [fms-hf-tuning](https://github.com/foundation-model-stack/fms-hf-tuning) library, which itself builds on the [`SFTTrainer` API from Hugging Face Transformers](https://huggingface.co/docs/trl/sft_trainer). The data is collected using AIM (AI Metadata) to capture profiling metadata during training runs.

- **Repository:** [ado SFTTrainer](https://github.com/IBM/ado/blob/main/plugins/actuators/sfttrainer/ado_actuators/sfttrainer/README.md)

## Uses

This dataset is intended to support the development of predictive models that map a proposed job configuration (model, hardware, method) to expected throughput and resource utilization — enabling "cost-calculator" or recommendation tools that guide users toward efficient fine-tuning setups. 

### Direct Use

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

See [ado Autoconf](https://github.com/IBM/ado/tree/main/plugins/custom_experiments/autoconf) for examples of how this data was used for building predictive models for automated configuration of resources for fine-tuning workloads

## Dataset Structure

These columns are produced from the data collected by the SFTTrainer actuator. The parameter descriptions have been sourced from the [ado SFTTrainer actuator documentation](https://ibm.github.io/ado/actuators/sft-trainer/#finetune_full_benchmark-v100). 

In case of any discrepancy, please refer to the actuator documentation for the most accurate and up-to-date information.
Key columns in the dataset:

| Column | Description |
| --- | --- |
| `model_name` | Name of the base LLM (e.g., `llama3-8b`, `mixtral-8x7b-instruct-v0.1`) |
| `method` | Fine-tuning technique used: `full`, `lora`, or `gptq-lora` |
| `number_gpus` | Number of GPUs allocated for the job |
| `gpu_model` | GPU model used (e.g., `L40S`) |
| `batch_size` | Number of samples processed per training step |
| `tokens_per_sample` | Sequence length of each sample |
| `is_valid` | Binary flag (1 or 0) indicating whether the run completed successfully |
| `dataset_tokens_per_second` | Primary throughput metric — tokens processed per second from the dataset; the most reliable performance measure |

Other columns capture additional metrics such as GPU memory utilization, GPU compute utilization, GPU power draw, and total training runtime.

## Dataset Creation

#### Data Collection and Processing

We benchmarked **30,190 jobs**, varying parameters of `SFTTrainer` including the base model, fine-tuning method, number of GPUs, and batch size. Training data consisted of random text, since the primary goal was to measure computational performance rather than model accuracy — the resulting fine-tuned model weights were discarded after each run.

## Citation [optional]

If you use this dataset, please cite it as:
```
@misc{lotito26finetuning,
  title={LLM Fine-Tuning Benchmark Dataset},
  author={Lotito, Daniele and Vassiliadis, Vassilis and  
          Pomponio, Alessandro and Venugopal, Srikumar and Johnston, Michael},
  howpublished={Hugging Face Datasets},
  url = {https://huggingface.co/datasets/ibm-research/LLMFineTuningBench/},
  year={2026}
}
```

## Contact Information

For any comments or questions, please email [Daniele Lotito](mailto:daniele.lotito@ibm.com), [Vassilis Vassiliadis](mailto:vassilis.vassiliadis@ibm.com) and [Srikumar Venugopal](mailto:srikumarv@ie.ibm.com).