README.md

---
language:
- en
license: gpl-3.0
tags:
- physics
- particle-physics
- lagrangian
annotations_creators:
- machine-generated
pretty_name: Particle Physics Lagrangian Dataset
source_datasets:
- none
task_categories:
- text2text-generation
configs:
- config_name: sampled
  data_files:
  - split: train
    path: ../datasets/huggingface_dataset_sampled.csv
  - split: eval
    path: ../datasets/huggingface_dataset_sampled.csv
- config_name: uniform
  data_files:
  - split: train
    path: ../datasets/huggingface_dataset_uniform.csv
  - split: eval
    path: ../datasets/huggingface_dataset_uniform.csv
dataset_info:
- config_name: sampled
  features:
  - name: fields
    dtype: string
  - name: Lagrangian
    dtype: string
  - name: train/eval
    dtype: string
  splits:
  - name: train
    num_examples: 228865
  - name: eval
    num_examples: 57217
- config_name: uniform
  features:
  - name: fields
    dtype: string
  - name: Lagrangian
    dtype: string
  - name: train/eval
    dtype: string
  splits:
  - name: train
    num_examples: 220552
  - name: eval
    num_examples: 55138
train-eval-index:
- config: sampled
  task: text2text-generation
  task_id: seq2seq
  splits:
    train_split: train
    eval_split: eval
  col_mapping:
    fields: fields
    Lagrangian: Lagrangian
    train/eval: train/eval
- config: uniform
  task: text2text-generation
  task_id: seq2seq
  splits:
    train_split: train
    eval_split: eval
  col_mapping:
    fields: fields
    Lagrangian: Lagrangian
    train/eval: train/eval
size_categories:
- 100K<n<1M
---

## Dataset Description

The **Particle Physics Lagrangian Dataset** was created to train a BART model for generating Lagrangians from particle fields and their symmetries. 
This task supports research in field theories within particle physics. Check [arXiv:2501.09729](http://https://arxiv.org/abs/2501.09729) for more information about the model.

### Data Generation

The dataset is generated through a pipeline utilizing AutoEFT, which helps automate the creation of effective field theories (EFTs). This tool is crucial for creating invariant terms based on specified fields and symmetries.

### Dataset Sampling

Due to the vast space of possible Lagrangians, careful sampling is essential:

1. **Uniform Dataset**: Provides evenly distributed Lagrangians for validation.
2. **Sampled Dataset**: Focuses on extreme cases to optimize learning, based on insights from natural language processing.

#### Key Features

- **Field Count**: Skews towards simpler Lagrangians with fewer fields.
- **Spin Types**: Includes a balanced mix of scalars and fermions.
- **Gauge Groups**: Uses SU(3), SU(2), and U(1) representations.
- **Trilinear Interaction Enrichment**: Includes crucial interaction terms fundamental to particle physics.

### Data Fields

- **fields**: List of input fields identified by their quantum numbers.
- **Lagrangian**: The corresponding Lagrangian for the input fields.
- **train/eval**: A flag describing whether the datapoint was used for training or evaluation.

### Encoding scheme

To facilitate understanding by the transformer model, the dataset undergoes a custom tokenization process that preserves the essential information of fields and Lagrangians:

- Fields and derivatives are tokenized to encapsulate quantum numbers, spins, and gauge symmetries.
- Key interactions are represented through positional tokens.
- Tokenization ensures all necessary contraction and symmetry details are conveyed.

For further details on the methods and theoretical underpinnings of this work, please refer to the paper "Generating Particle Physics Lagrangians with Transformers" [arXiv:xxxx.xxxxx](https://arxiv.org/abs/xxxx.xxxxx). 


### Usage

Ideal for sequence-to-sequence tasks, this dataset is optimized for training transformer models to derive particle physics Lagrangians efficiently.