Datasets:

Zovda
/

dm0019-ecommerce-attribute-extraction-sft-samples

Structured SFT corpus for e-commerce attribute extraction: every record carries instruction, input, output and field-level evidence chains, in full-text and title-only forms — 2,808 records, v2.1.1 signed off after passing 39/39 QC checks.

面向电商商品属性抽取的结构化 SFT 语料：每条记录含指令、输入、输出与字段级证据链，覆盖「全文（标题+详情）」与「仅标题」两种形态，2,808 条记录，v2.1.1 签版质检 39/39 全过。

What's inside / 数据内容

main: full-text extraction from title + detail, 1,485 records (1,170 with brand).
main 形态：基于标题 + 详情文本的全文抽取，1,485 条（含品牌 1,170 条）。
title_only: only attributes with title evidence, no guessing — 1,323 records (789 with brand), paired with main.
title_only 形态：仅输出标题中有证据的属性、不臆测，1,323 条（含品牌 789 条），与 main 同源成对。
evidence_hit logs, per field, the raw span → normalised value / trigger word → hit result, plus the drop reason for discarded attributes (no_evidence / junk_value / noinfo_value).
evidence_hit 逐字段记录 raw 原文 → 规范值 / 驱动词 → 命中结果，并记录被丢弃属性的原因（no_evidence / junk_value / noinfo_value）。
Archive files are delivered on request — contact lucas@zovda.com. This page documents the full schema with real examples.
样例归档文件按需交付，请联系 lucas@zovda.com；本页给出完整字段结构与真实样例。

Specifications / 规格

Record schema v2: instruction / input / output / evidence_hit / meta
记录 schema v2：instruction / input / output / evidence_hit / meta
Output is normalised attribute JSON: keys are canonical attribute names, values are canonical values
输出为规范属性 JSON：键为标准属性名，值为规范值
Version: v2.1.1 (signed off 2026-09-18, frozen baseline)
版本：v2.1.1（2026-09-18 签版，冻结基线）

Suitable tasks / 适用任务

Attribute extraction and listing-information extraction
属性抽取与商品刊登上架信息抽取
Evidence-constrained output: emit only when evidence exists, suppressing hallucination
证据约束输出：仅在有证据时输出，抑制幻觉
Title-only vs full-text information-boundary training and evaluation
「仅标题 vs 全文」信息边界对照训练与评测
Product data governance and consistency review
商品数据治理与一致性审核

Files / 文件清单

本数据集为说明页：样例文件按需交付，请联系 lucas@zovda.com 获取。

This card is documentation-only: sample files are delivered on request — contact lucas@zovda.com.

Usage / 使用方式

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Zovda/dm0019-ecommerce-attribute-extraction-sft-samples",
    repo_type="dataset",
    local_dir="./dm0019-ecommerce-attribute-extraction-sft-samples",
)

Rights and disclaimer / 授权与声明

Text derives from public product-page snapshots and remains the property of their rightsholders. Samples demonstrate schema and annotation conventions; contact us for commercial licensing.

文本来源于公开商品详情页快照，著作权归相应权利人。样例用于展示字段结构与标注口径；商用授权请联系我们。

This dataset is shared to demonstrate data organisation, technical specifications and delivery structure. For formal licensing, large-scale access or custom production, contact lucas@zovda.com.

本数据集仅用于展示数据组织方式、技术规格与交付结构。正式授权、规模化取数与定制生产请联系 lucas@zovda.com。

About Zovda / 关于 Zovda

Zovda builds, cleans, annotates and packages training data across video, audio, images, e-commerce image-text pairs, exam question banks, embodied AI and coding-agent corpora.

Zovda 提供多模态与文本训练语料的数据采集、清洗、标注与结构化交付服务，覆盖视频、音频、图像、电商图文、K12 与职业题库、具身智能、代码智能体等方向。

Contact / 联系方式: lucas@zovda.com

Downloads last month: 87

Total file size:

7.57 kB