File size: 5,324 Bytes
5e4510c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
"""
Unified data loading interface for scaling law discovery.

Dynamically loads data from the Hugging Face Hub repository 'pkuHaowei/sldbench'.
This approach centralizes data access and ensures consistency.
"""
import numpy as np
import datasets
from typing import Dict, Any, Tuple

# --- Configuration ---

HUB_REPO_ID = "pkuHaowei/sldbench"

# Defines the schema for each task, mapping feature/target names from the Hub
# to the columns in the dataset.
TASK_SCHEMA_MAP = {
    "data_constrained_scaling_law": {
        "feature_names": ["unique_tokens", "params", "tokens"],
        "target_name": "loss",
    },
    "domain_mixture_scaling_law": {
        "feature_names": [f"proportion_domain_{i+1}" for i in range(5)],
        "target_name": [f"loss_domain_{i+1}" for i in range(5)],
    },
    "lr_bsz_scaling_law": {
        "feature_names": ["lr", "bsz", "data_size", "non_embedding_param_size"],
        "target_name": "lm_loss",
    },
    "moe_scaling_law": {
        "feature_names": ["num_experts", "dense_parameter_count"],
        "target_name": "loss_validation",
    },
    "sft_scaling_law": {
        "feature_names": ["sft_data_size"],
        "target_name": "sft_loss",
    },
    "vocab_scaling_law": {
        "feature_names": ["non_vocab_parameters", "vocab_size", "num_characters"],
        "target_name": "unigram_normalized_loss",
    },
    "parallel_scaling_law": {
        "feature_names": ["num_params", "parallel_size"],
        "target_name": "loss"
    },
    "easy_question_scaling_law": {
        "feature_names": ["log_flops"],
        "target_name": "brier_score",
    }
}

def load_data(
    app_name: str,
    train: bool = True,
) -> Dict[Any, Tuple[np.ndarray, np.ndarray]]:
    """
    Unified data loading interface. Loads and processes data from Hugging Face Hub.

    Each task's dataset is grouped by a 'group' key. The function returns a 
    dictionary mapping each group key to a tuple of (features, labels).
    - features (X): A numpy array of shape (n_samples, n_features).
    - labels (y): A numpy array of shape (n_samples,) or (n_samples, n_targets).

    Args:
        app_name: The name of the task (e.g., 'sft_scaling_law').
        train: If True, load training data; otherwise, load test data.

    Returns:
        A dictionary containing the prepared data, structured by group.
    """
    if app_name not in TASK_SCHEMA_MAP:
        raise ValueError(f"Task '{app_name}' not found in TASK_SCHEMA_MAP. Available tasks: {list(TASK_SCHEMA_MAP.keys())}")

    split = 'train' if train else 'test'
    schema = TASK_SCHEMA_MAP[app_name]
    
    try:
        # Load the specific task dataset from the Hugging Face Hub
        dataset = datasets.load_dataset(HUB_REPO_ID, name=app_name, split=split)
    except Exception as e:
        raise IOError(f"Failed to load dataset '{app_name}' with split '{split}' from '{HUB_REPO_ID}'. Reason: {e}")

    # Ensure target_name is a list for consistent processing
    feature_names = schema["feature_names"]
    target_names = schema["target_name"]
    if not isinstance(target_names, list):
        target_names = [target_names]

    processed_data = {}
    
    # The datasets are partitioned by a 'group' column
    unique_groups = sorted(list(set(dataset['group'])))

    for group_key in unique_groups:
        # Filter the dataset for the current group
        group_data = dataset.filter(lambda example: example['group'] == group_key)
        
        # Extract features (X) and stack them into a single numpy array
        X_list = [np.array(group_data[fname]) for fname in feature_names]
        X = np.stack(X_list, axis=1)
        
        # Extract targets (y)
        y_list = [np.array(group_data[tname]) for tname in target_names]
        y_stacked = np.stack(y_list, axis=1)
        
        # Squeeze the last dimension if there is only one target
        y = y_stacked.squeeze(axis=1) if y_stacked.shape[1] == 1 else y_stacked
        
        processed_data[group_key] = (X, y)

    return processed_data

if __name__ == '__main__':
    # Example of how to use the new loader
    # The list of tasks is now derived directly from the schema map
    ALL_TASKS = list(TASK_SCHEMA_MAP.keys())

    for task in ALL_TASKS:
        print(f"\n--- Testing '{task}' ---")
        try:
            # Load training data
            train_data = load_data(task, train=True)
            print(f"Successfully loaded training data from Hugging Face repo '{HUB_REPO_ID}'.")
            
            # Inspect the first group's shape
            first_group_key = next(iter(train_data))
            X_train, y_train = train_data[first_group_key]
            print(f"Train groups: {len(train_data)}. First group '{first_group_key}' shape: X={X_train.shape}, y={y_train.shape}")
            
            # Load test data
            test_data = load_data(task, train=False)
            if test_data:
                first_test_key = next(iter(test_data))
                X_test, y_test = test_data[first_test_key]
                print(f"Test groups: {len(test_data)}. First group '{first_test_key}' shape: X={X_test.shape}, y={y_test.shape}")
            else:
                print("Test data is empty.")

        except (ValueError, IOError, KeyError) as e:
            print(f"Error loading data for task '{task}': {e}")