permutation-groups.py

import datasets
import json
import os
import pandas as pd

_DESCRIPTION = "Permutation composition datasets with dynamic filtering by group degree, order, and sequence length."
_HOMEPAGE = "https://huggingface.co/datasets/BeeGass/permutation-groups"
_LICENSE = "MIT"

class PermutationGroupsConfig(datasets.BuilderConfig):
    def __init__(
        self, 
        group_type=None,
        min_degree=None,
        max_degree=None,
        min_order=None,
        max_order=None,
        min_len=3,
        max_len=1024,
        **kwargs
    ):
        """
        Configuration for loading permutation groups.
        
        Args:
            group_type: Type of group (symmetric, alternating, cyclic, dihedral, klein, 
                       quaternion, elementary_abelian, psl, frobenius, mathieu)
            min_degree: Minimum group degree to include
            max_degree: Maximum group degree to include
            min_order: Minimum group order to include
            max_order: Maximum group order to include
            min_len: Minimum sequence length
            max_len: Maximum sequence length
        """
        # Set name based on parameters
        if "name" not in kwargs:
            if group_type:
                kwargs["name"] = group_type
            else:
                kwargs["name"] = "all"
        
        super().__init__(**kwargs)
        self.group_type = group_type
        self.min_degree = min_degree
        self.max_degree = max_degree
        self.min_order = min_order
        self.max_order = max_order
        self.min_len = min_len
        self.max_len = max_len

class PermutationGroups(datasets.GeneratorBasedBuilder):
    """Permutation groups dataset with dynamic filtering."""
    
    VERSION = datasets.Version("5.0.0")
    
    # Define all available group types
    GROUP_TYPES = [
        "symmetric", "alternating", "cyclic", "dihedral", 
        "klein", "quaternion", "elementary_abelian", "psl", 
        "frobenius", "mathieu"
    ]
    
    BUILDER_CONFIGS = []
    
    # Add configs for each group type
    for group_type in GROUP_TYPES:
        BUILDER_CONFIGS.append(
            PermutationGroupsConfig(
                name=group_type,
                description=f"{group_type.capitalize()} permutation groups",
                group_type=group_type,
            )
        )
    
    # Add "all" configuration
    BUILDER_CONFIGS.append(
        PermutationGroupsConfig(
            name="all",
            description="All permutation groups",
            group_type=None,  # Will load all types
        )
    )
    
    DEFAULT_CONFIG_NAME = "symmetric"
    
    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features({
                "input_sequence": datasets.Value("string"),
                "target": datasets.Value("string"),
                "group_type": datasets.Value("string"),
                "group_degree": datasets.Value("int32"),
                "group_order": datasets.Value("int32"),
                "sequence_length": datasets.Value("int32"),
            }),
            homepage=_HOMEPAGE,
            license=_LICENSE,
        )
    
    def _split_generators(self, dl_manager):
        # Determine which datasets to load
        if self.config.group_type:
            # Load the superset for this group type
            datasets_to_load = [f"{self.config.group_type}_superset"]
        else:
            # Load all supersets
            datasets_to_load = [
                "symmetric_superset", "alternating_superset", 
                "cyclic_superset", "dihedral_superset",
                "klein_superset", "quaternion_superset",
                "elementary_abelian_superset", "psl_superset",
                "frobenius_superset", "mathieu_superset"
            ]
        
        # Build file URLs using wildcards
        train_urls = []
        test_urls = []
        
        for dataset_name in datasets_to_load:
            train_urls.append(f"data/{dataset_name}/train/data-*.arrow")
            test_urls.append(f"data/{dataset_name}/test/data-*.arrow")
        
        # Download files
        downloaded_files = dl_manager.download({
            "train": train_urls,
            "test": test_urls
        })
        
        # Flatten the lists of files
        train_files = []
        test_files = []
        
        for file_list in downloaded_files["train"]:
            if isinstance(file_list, list):
                train_files.extend(file_list)
            else:
                train_files.append(file_list)
                
        for file_list in downloaded_files["test"]:
            if isinstance(file_list, list):
                test_files.extend(file_list)
            else:
                test_files.append(file_list)
        
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "files": train_files,
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "files": test_files,
                },
            ),
        ]
    
    def _generate_examples(self, files):
        """Yield examples with filtering."""
        idx = 0
        
        for file_path in files:
            # Load the Arrow file
            table = datasets.table.read_table(file_path)
            
            # Convert to pandas for easier filtering
            df = table.to_pandas()
            
            # Apply filters
            mask = pd.Series([True] * len(df))
            
            # Filter by group type (if specified in config)
            if self.config.group_type:
                mask &= (df["group_type"] == self.config.group_type)
            
            # Filter by degree
            if self.config.min_degree is not None:
                mask &= (df["group_degree"] >= self.config.min_degree)
            if self.config.max_degree is not None:
                mask &= (df["group_degree"] <= self.config.max_degree)
            
            # Filter by order
            if self.config.min_order is not None:
                mask &= (df["group_order"] >= self.config.min_order)
            if self.config.max_order is not None:
                mask &= (df["group_order"] <= self.config.max_order)
            
            # Filter by sequence length
            if self.config.min_len is not None:
                mask &= (df["sequence_length"] >= self.config.min_len)
            if self.config.max_len is not None:
                mask &= (df["sequence_length"] <= self.config.max_len)
            
            # Apply mask
            filtered_df = df[mask]
            
            # Yield filtered examples
            for _, row in filtered_df.iterrows():
                yield idx, {
                    "input_sequence": row["input_sequence"],
                    "target": row["target"],
                    "group_type": row["group_type"],
                    "group_degree": int(row["group_degree"]),
                    "group_order": int(row["group_order"]),
                    "sequence_length": int(row["sequence_length"]),
                }
                idx += 1