Field ExtractionA-layerHard facts only

搜推广字段化抽取

把 A 层资源进一步拆成可机器读取的字段：已抽事实、缺失字段和下一步全文/产品页精读任务。

5resources

40fields

A 层资源字段表

只抽当前 hard evidence 支持的字段；缺全文的保留 missing_fields。

partial_from_abstract

Yambda-5B

Evidence: src_yambda_5b

Yambda-5B is a benchmark-grade hard asset because it exposes scale, feedback types, modality and temporal evaluation protocol, not only platform existence.

asset_type	large-scale open dataset for ranking and retrieval
source_platform	Yandex Music streaming platform
scale_users	1 million users
scale_items	9.39 million tracks
scale_interactions	4.79 billion user-item interactions
implicit_feedback	listening events
explicit_feedback	likes, dislikes, unlikes, undislikes
modalities	audio embeddings for most tracks generated from audio spectrograms
key_flag	is_organic separates organic user actions from recommendation-driven events
evaluation_protocol	Global Temporal Split
baselines_or_models	ItemKNN iALS SANSA SASRec

Missing:

licensedownload URLfield dictionaryexact metricsleaderboard or code repositoryfull PDF/HTML reading

Next: Recover full text and dataset card, then split fields into scale table, schema table and benchmark task table.

partial_from_abstract

VK-LSVD

Evidence: src_vk_lsvd

VK-LSVD turns VK into a hard recommendation benchmark resource because it exposes industrial-scale interaction and feature structure.

asset_type	industrial short-video recommendation dataset
scale_users	10 million users
scale_items	almost 20 million videos
scale_interactions	over 40 billion interactions
time_span	six months
features	content embeddings diverse feedback signals contextual metadata
challenge_or_usage	central role in VK RecSys Challenge 2025
target_tasks	sequential recommendation cold-start scenarios next-generation recommender systems

Missing:

dataset licensedownload URLfull field schemabaseline tablemetricschallenge rulesfull paper reading

Next: Recover full paper/challenge page and create task-schema-metric extraction.

partial_from_abstract

Yandex personalized commerce ranking

Evidence: src_yandex_ecommerce_ranking

This is the strongest Yandex ranking-method evidence because it links model architecture, training, debiasing, cross-signal use and deployment into one resource.

problem_context	commerce homepage and other cases where user intent is not explicit
core_task	personalized ranking for e-commerce recommendations
model_idea	transformer-based two-tower model architecture
training_idea	two-stage training process fine-tunes two-tower models for ranking performance
retrieval_to_ranking_gap	features from retrieval-focused deep learning models are sub-optimal for ranking stage
debiasing	novel technique for debiasing context in offline models
cross_signal	uses web-search queries for e-commerce recommendations
deployment	deployed at Yandex and serves millions of users daily
evaluation	reports strong performance in online A/B testing

Missing:

offline metric namesA/B metric namesarchitecture detailsfeature setloss functionsbaseline valuesfull paper reading

Next: Recover full paper and extract model diagram, task setup, offline and online metric tables.

partial_from_official_snapshot

2GIS local discovery promotion

Evidence: src_2gis_ads_candidate

2GIS provides hard official evidence for a local search-promo layer: merchant visibility is tied to location, directory and action conversion rather than only web query ranking.

asset_type	official local business promotion page
surface	2GIS map/directory/local discovery surface
business_goal	attract attention to business
visible_sections	cases advertising options more about advertising request form
contact_channel	phone and request form are visible in snapshot
regional_scope_seen	Russia Armenia Belarus Georgia Kazakhstan Kyrgyzstan Tajikistan Uzbekistan
intent_layer	local merchant and city-level discovery intent

Missing:

ad formatspricingmerchant workflowcase detailscity coverageconversion metrics

Next: Split the official snapshot into product-format, city-coverage and merchant-workflow subcards.

partial_from_official_snapshot_and_cdp

WB seller promotion instructions

Evidence: extra_wb_promo_instructions

WB evidence is a product-process anchor: it shows seller-side promotion operations and measurable workflow topics, not just a marketplace overview.

asset_type	official seller instruction page
surface	seller portal training/help center
visible_topics	promotion payment models and operating principles launch and edit promotion campaign choose promotion bid budget and payment promotion statistics and management FAQ about search and promotion
cdp_support	browser/CDP observed a full seller promotion instructions page with 19119 text characters
intent_layer	marketplace seller-side traffic and exposure operation

Missing:

promotion models detailsbid rulesbudget formulastatistics fieldsquality controlssection-level translation/extraction

Next: Parse the CDP text by visible topic and produce seller workflow, bidding, budget and statistics extraction tables.