TAAC-25 / meta_data.json
salmon1802's picture
Upload folder using huggingface_hub
a23f0b6 verified
Raw
History Blame Contribute Delete
6.55 kB
{
"sample_size": {
"total": 757207146,
"train": 586117494,
"valid": 17846140,
"test": 153243512,
"phase1_total_interactions": 976291071
},
"split_by_minimal_tail_with_positive_constraints": {
"timezone": "Asia/Shanghai",
"timestamp_unit": "s",
"train_days": 226,
"train_range": "2024-10-15 ~ 2025-05-28",
"valid_days": 1,
"valid_range": "2025-05-29 ~ 2025-05-29",
"test_days": 68,
"test_range": "2025-05-30 ~ 2025-08-05",
"valid_click_pos": 841303,
"valid_conversion_pos": 396010,
"test_click_pos": 553866,
"test_conversion_pos": 250843,
"rule": "test=最后最短合法正样本时间段; valid=紧邻test的最短合法正样本时间段; 其余=train"
},
"user_filtering": {
"min_user_interactions": 10,
"dropped_all_negative_users": 2424604,
"rule": "用户完整行为序列若既无 click 也无 conversion,则过滤"
},
"oov_filter": {
"min_feat_count": 2,
"applied_to": {
"scalar_features": [
"103",
"104",
"105",
"109",
"100",
"101",
"102",
"112",
"114",
"115",
"116",
"117",
"118",
"119",
"120",
"121",
"122"
],
"list_features": [
"106",
"107",
"108",
"110"
]
},
"rule": "特征值出现次数 < min_feat_count 的统一映射为 0 (unknown/padding)",
"freq_source": "基于日志中 user/item 出现次数 * 特征值计算"
},
"blocked_layout": {
"train_blocks": 16,
"valid_blocks": 8,
"test_blocks": 8,
"train": {
"data_dir": "train/data",
"user_info_dir": "train/user_info",
"item_info_dir": "train/item_info"
},
"valid": {
"data_dir": "valid/data",
"user_info_dir": "valid/user_info",
"item_info_dir": "valid/item_info"
},
"test": {
"data_dir": "test/data",
"user_info_dir": "test/user_info",
"item_info_dir": "test/item_info"
},
"block_pair_rule": "同一 split 下,data/user_info/item_info 使用相同的 part-xxxxx 编号配对读取。",
"local_index_rule": {
"user_index": "block-local dense index, starts from 0",
"item_index": "block-local dense index, 0 reserved for padding",
"user_id": "global feature id, consistent across blocks",
"item_id": "global feature id, consistent across blocks"
}
},
"vocab_size": {
"user_id": 7706779,
"item_id": 17487677,
"timestamp": 0,
"action": 4,
"103": 88,
"104": 3,
"105": 8,
"109": 4,
"106": 15,
"107": 20,
"108": 5,
"110": 3,
"100": 7,
"101": 54,
"102": 173463,
"112": 31,
"114": 34,
"115": 989,
"116": 21,
"117": 559,
"118": 1637,
"119": 4951,
"120": 4046,
"121": 5041300,
"122": 2389,
"day_of_week": 8,
"is_weekend": 3,
"hour": 25
},
"label": [
"is_click",
"is_conversion"
],
"action_vocab": {
"click": 1,
"conversion": 2,
"exposure": 3
},
"action_vocab_desc": "编码后的 action 词表,用于 dataloader 基于 full_action_seq 构造 task-specific token masks。",
"action_mapping": {
"raw_action_type_to_name": {
"0": "exposure",
"1": "click",
"2": "conversion"
},
"action_name_to_code": {
"click": 1,
"conversion": 2,
"exposure": 3
}
},
"user_info_schema": {
"fields": [
"user_index",
"full_item_seq",
"full_action_seq",
"full_timestamp_seq"
],
"desc": "这里的 user_index / full_item_seq 中的 item index 都是 block-local index;full_action_seq / full_timestamp_seq 为全局时间顺序序列。"
},
"item_info_schema": {
"fields": [
"item_index",
"item_id",
"100",
"101",
"102",
"112",
"114",
"115",
"116",
"117",
"118",
"119",
"120",
"121",
"122"
],
"desc": "item_index 为 block-local index;item_id 为全局 item feature id,用于 embedding 一致性。"
},
"feature_schema": {
"user_scalar_features": [
"103",
"104",
"105",
"109"
],
"user_list_features": [
"106",
"107",
"108",
"110"
],
"user_list_keep": 5,
"user_list_storage": "single-column list[int] with fixed length",
"user_scalar_encoding": "keep raw integer ids; vocab_size=max+1; OOV filtered",
"user_list_encoding": "keep raw list[int]; vocab_size=max_sub_feature_id+1; OOV filtered",
"item_static_features": [
"100",
"101",
"102",
"112",
"114",
"115",
"116",
"117",
"118",
"119",
"120",
"121",
"122"
],
"item_static_encoding": "keep raw integer ids; vocab_size=max+1; OOV filtered",
"context_features": [
"day_of_week",
"is_weekend",
"hour"
],
"time_feature_encoding": {
"day_of_week": "0~6 -> 1~7, 0 reserved for padding/unknown",
"is_weekend": "0/1 -> 1~2, 0 reserved for padding/unknown",
"hour": "0~23 -> 1~24, 0 reserved for padding/unknown"
},
"timestamp_kept_raw": true,
"context_time_timezone": "Asia/Shanghai"
},
"max_len": {
"full_item_seq": 100,
"full_action_seq": 100,
"full_timestamp_seq": 100
},
"raw_action_type_counter": {
"0": 923826708,
"1": 27813405,
"2": 24650958
}
}