SearchJev/search_jev_unified_v2_release_20260928

SearchJev Training Data

The official training data of SearchJev: A Fast and Calibrated System-1 Model for Search Agents.

Paper · Code · Models · Citation

Language / 语言： English | 中文

---

English

Overview

This dataset is the training data used to build SearchJev, a discriminative System-1 model that makes fast, calibrated judgments for search agents. Instead of generating free text, SearchJev reads a state (a query, a search-results page, retrieved evidence, or a partial agent trajectory) together with a set of typed questions, and returns a structured answer with a probability distribution for each question. Typical questions include whether to keep searching or answer now, which result is worth opening, whether the evidence is sufficient, which follow-up query to issue, and whether a document is relevant, timely and authoritative.

The dataset unifies 73 Chinese and English sources (public benchmarks, search-agent trajectories, mined hard negatives and purpose-built synthetic scenarios) into a single format, with group-aware splits and per-question loss weights for multi-task training.

• Paper: https://arxiv.org/abs/2610.05107
• Code: https://github.com/EvoScientist/SearchJev
• Models: https://huggingface.co/collections/SearchJev/searchjev
• Release version: v3.0.0 (2026-10-02). The directory name search_jev_unified_v2_release_20260928 is kept for compatibility; the authoritative version is release_version in manifest.json.

Highlights

• Search-decision focused. 38 capabilities organised into 8 decision buckets, from next-action and stop-search decisions to query–document relevance, evidence verification and query understanding.
• One schema for every task. Each row is a state plus one or more typed questions (noul / choice / score) with reference answers, so heterogeneous tasks can be trained and evaluated jointly.
• Leakage-controlled splits. Rows that share a mother sample, a translation or mining parent, or an identical input are grouped and assigned to a single split (test > val > train). Cross-split groups and exact duplicate inputs are both zero.
• Training-ready metadata. A line-aligned index file carries per-question loss weights, evaluation eligibility, label provenance, decision bucket and a full-state token pre-check (every retained question fits in 6,144 tokens with the Qwen3 tokenizer).
• Bilingual. About 64.5% of training rows are Chinese-dominant and 35.4% are Latin-script-dominant.

Statistics

Split	Rows	Questions
train	467,368	646,664
val	138,321	172,221
test	189,708	229,084
Total	795,397	1,047,969

Additional evaluation resources:

Resource	Size	Purpose
Model-selection set	5,000 rows	Checkpoint selection; covers 66 validation source identities and all 38 capabilities, one row per mother group
Calibration set	2,000 rows	Reserved for probability calibration (temperature / thresholds); disjoint from the model-selection set and test groups

Task Taxonomy

Every row belongs to one decision bucket. The loss share is the recommended share of total training loss, implemented by the per-bucket multipliers in training_config.json.

Bucket	Description	Train rows	Loss share
A	Search actions: next action, stop vs. continue, marginal gain, next-hop query selection	120,025	40%
B1	Query–document quality: timeliness, authority, user satisfaction	61,221	10%
B2	Query–document relevance and ranking	55,376	12%
C	Evidence and fact verification	41,288	12%
D1	Sentence-pair and rewrite equivalence	65,465	8%
D2	Query assessment: search need, information content, noise	42,729	8%
D3	Query intent classification	29,737	4%
E	General understanding: sentiment, NLI, topic, preference	51,527	6%

The 38 capabilities are: action_imitation, clarify_decision, contextual_rewrite, coreference, emotion, evidence, evidence_sufficiency, evidence_sufficiency_state, failure_recovery, fresh_serp, intent, intent_equivalence, keyword_match, logic_mc, marginal_gain, next_query, next_query_zh, nli, preference_pair, qd_quality, query_classification, query_fit, query_info, query_need, query_noise, reading_mc, relevance, response_quality, retrieval_gain, review_rating, rewrite, satisfaction, sentiment_binary, serp_select, source_time, state_action, stop_boundary, topic_classification.

Synthetic Search-Decision Scenarios

To cover decisions that public datasets rarely label, the dataset includes 27,183 rows of purpose-built synthetic scenarios (input_origin = synth_v1) and 10,109 rows of multi-dimensional quality data (synth_tas).

Track	Scenario	train	val	test
T1	State-conditioned action: evidence sufficiency, conflict, budget, next step	2,710	165	185
T2	SERP selection: which results are worth opening	494	28	32
T3	Marginal gain: new information, redundant evidence, conflict, freshness	1,802	117	124
T4	Next-hop query selection (Chinese candidates)	3,663	440	366
T5	Source and time: authority and timeliness	3,762	265	293
T6	Failure recovery: failed searches or page reads, unusable results, injected content	5,297	313	366
T7	Contextual rewrite: intent preservation in multi-turn dialogue	1,320	317	271
T10	Stop boundary: sufficient, outdated, conflicting and redundant evidence	378	28	20
T11	Clarification: when to ask the user and what to do next	3,744	374	309

See SYNTHETIC_SCENARIOS.md for details.

Data Sources

Family	Representative sources
Retrieval and relevance	T2Ranking, DuRetrieval, Multi-CPR (e-commerce / medical / video), QBQTC, KUAKE-QQR, LCQMC, BQ Corpus, ATEC, AFQMC, CHIP-STS, PAWS-X (zh), QReCC, dialogue rewrite equivalence, query–document quality annotations
Evidence and verification	VitaminC, CFEVER, TrendFact, RAGTruth
Search decisions	Search-agent trajectories, KUAKE-QIC, Search Arena, AgentRewardBench
Search hard negatives	Mined hard negatives for gain, next hop, query fit and sufficiency, including Chinese translations and a harder subset
Query understanding	CLINC150, Banking77, MASSIVE (zh), query information / search-need / noise assessment
General understanding	CLUE (CMNLI, OCNLI, TNEWS, CSL, WSC, C3), LogiQA (zh), ChnSentiCorp, Weibo Senti 100k, Online Shopping 10 Cats, THUCNews titles, Amazon Reviews Multi (zh), Douban and JD reviews, multi-turn emotion dialogue, COIG-P, CValues, HelpSteer2 / HelpSteer3
Synthetic	Search-decision scenarios T1–T11 and multi-dimensional quality data (see above)

Some sources (e.g. Search Arena, AgentRewardBench, RAGTruth, the QReCC test split) appear only in the test split and serve as held-out evaluation.

Data Format

Each line of {train,val,test}.jsonl is one sample: a state, a dictionary of questions, the reference answers, and a difficulty level.

{
  "state": {
    "premise": "我走到近前,见是一位工人模样的中年妇女",
    "hypothesis": "我看见了一位妇女。"
  },
  "questions": {
    "nli": {
      "type": "choice",
      "instructions": "前提能不能撑起假设？还是矛盾或无关？",
      "criteria": {
        "entailment": "蕴含（可由前提推出）",
        "neutral": "无关（前提不能确定）",
        "contradiction": "矛盾（与前提冲突）"
      }
    }
  },
  "answers": {
    "nli": {
      "type": "choice",
      "choice": "entailment",
      "confidence": 1.0,
      "probabilities": {"entailment": 1.0, "neutral": 0.0, "contradiction": 0.0}
    }
  },
  "difficulty": 3
}

Question types

Type	Meaning	criteria	Answer
noul	Is the proposition true?	optional {"true": ..., "false": ...}	noul: probability of true
choice	Pick one option from a fixed set	{option: description}	choice, confidence, probabilities over all options
score	Ordinal rating with 2–10 levels	list of level descriptions, low to high	score (expected level), confidence, probabilities keyed by level

state may be a string, a JSON object or a list. A row may contain several questions about the same state; each question is answered independently.

Index files. {split}_index.jsonl is aligned line by line with {split}.jsonl and holds metadata. Key fields:

Field	Description
sample_id, source, input_origin	Sample identifier and provenance
capability, task_family, decision_bucket	Task taxonomy
training_lane, supervision, weak, label_tier, target_provenance	Label provenance and supervision strength
loss_weight, row_loss_scale	Per-question loss weights and row-level scale
evaluation_eligible, eval_track	Whether each question is used in evaluation, and its evaluation track
gold, confidence, teacher_reference_probability	Reference answers and auxiliary teacher probabilities
canonical_group, group_keys	Grouping keys used for leakage-free splitting
language_heuristic, difficulty, token_preflight	Language, difficulty and token-length statistics

Files

File	Description
train.jsonl, val.jsonl, test.jsonl	Main data
train_index.jsonl, val_index.jsonl, test_index.jsonl	Line-aligned metadata
model_selection.jsonl, model_selection_index.jsonl	Fixed checkpoint-selection set
calibration_blind.jsonl, calibration_reference.jsonl	Calibration set: questions without answers, and the corresponding reference labels
training_config.json	Reference training configuration (Qwen3-1.7B, LoRA, bucket multipliers)
budget.json	Loss budget by bucket
manifest.json	Release metadata and SHA-256 checksums
data_report.json, DATA_REPORT.md	Statistics and detailed construction report
SYNTHETIC_SCENARIOS.md	Synthetic scenario documentation
VERSION_HISTORY.md, VERSION_HISTORY.json	Version history
changes.jsonl	Per-row change log relative to the previous release

Usage

import json

with open("train.jsonl", encoding="utf-8") as f_data, \
     open("train_index.jsonl", encoding="utf-8") as f_index:
    for line, meta_line in zip(f_data, f_index):
        sample, meta = json.loads(line), json.loads(meta_line)
        for qid, question in sample["questions"].items():
            answer = sample["answers"][qid]
            weight = meta["loss_weight"][qid] * meta["row_loss_scale"]
            ...

For training, we recommend one question per forward pass (independent context), weighting each question by loss_weight × row_loss_scale × bucket multiplier. The SearchJev repository provides a reference training and evaluation pipeline that reads this directory directly.

Construction and Quality Control

• Unified schema. All sources are mapped to the state / typed-question / answer format; instructions are diversified to prevent template overfitting.
• Label cleaning. Questions with suspected or disputed labels, unadjudicated negatives, unresolved rubric mappings, or shortcut features (e.g. future actions visible in the input) are removed from supervision rather than relabelled automatically.
• Leakage control. Group-aware splitting over mother samples, translation and mining parents, and identical inputs; exact duplicate inputs are removed.
• Length control. Every retained question is checked against a 6,144-token budget with its full state. Over-length questions are removed as a whole instead of being truncated.
• Loss budgeting. Bucket multipliers are computed so the realised loss shares match the target A/B1/B2/C/D1/D2/D3/E = 40/10/12/12/8/8/4/6%.

Limitations

• Targets are hard reference labels derived from the original annotations, demonstrations or construction rules. The confidence = 1.0 in answers encodes the target and is not an empirical accuracy; calibrated probabilities should be fitted on the calibration set.
• Labels in the action-imitation track describe what a demonstrator did, not necessarily the optimal policy.
• Bucket A is still largely built from derived English multi-hop data; Chinese rows make up about 19% of bucket A in train.
• The model-selection set is stratified and should not be read as an estimate of accuracy on the natural distribution.

License

This dataset is derived from multiple public datasets and synthetic data. Each subset remains subject to the license and terms of use of its original source. Please consult the upstream licenses before use. The dataset is intended for research purposes.

Citation

If you use this dataset, please cite:

@misc{cao2026searchjev,
  title  = {{SearchJev}: A Fast and Calibrated System-1 Model for Search Agents},
  author = {Cao, Congfeng and Zuo, Lipeng and Papakostas, Konstantinos and Xu, Qiwei and Xu, Songwei and
            Zhou, Lun and Ren, Zhaochun and Lyu, Yougang and Yan, Xiaohui},
  year   = {2026}
}

---

中文

概述

本数据集是 SearchJev 的官方训练数据。SearchJev 是面向搜索智能体的快速、可校准的 System-1 判别模型：它不生成自由文本，而是读取一份状态（查询、搜索结果页、检索证据或智能体的部分轨迹）和一组类型化问题，为每个问题输出结构化答案及概率分布。典型问题包括：继续搜索还是直接作答、哪条结果值得打开、证据是否充分、下一跳应发起什么查询、文档是否相关、及时和权威等。

数据集将 73 个中英文来源（公开基准、搜索智能体轨迹、挖掘的难负例以及专门构造的合成场景）统一为同一格式，采用按组切分，并提供题级损失权重，可直接用于多任务训练。

• 论文： https://arxiv.org/abs/2610.05107
• 代码： https://github.com/EvoScientist/SearchJev
• 模型： https://huggingface.co/collections/SearchJev/searchjev
• 发布版本： v3.0.0（2026-10-02）。目录名 search_jev_unified_v2_release_20260928 为兼容而保留，版本以 manifest.json 中的 release_version 为准。

主要特点

• 聚焦搜索决策。 38 项能力归入 8 个决策桶，覆盖下一步动作与停止搜索决策、查询–文档相关性、证据核查和查询理解等。
• 统一格式。 每行由一个 state 和一道或多道类型化问题（noul / choice / score）及参考答案组成，异构任务可联合训练与评测。
• 防泄漏切分。 共享母本、翻译或挖掘父本、相同输入的样本归为一组，整组划入同一切分（优先级 test > val > train）。跨切分组与完全重复输入均为 0。
• 开箱即用的元数据。 逐行对齐的索引文件提供题级损失权重、评测资格、标签来源、决策桶及完整状态长度预检（全部保留题在 Qwen3 分词器下不超过 6,144 token）。
• 中英双语。 训练集中约 64.5% 的行以中文为主，35.4% 以拉丁字母文本为主。

数据规模

切分	行数	题数
train	467,368	646,664
val	138,321	172,221
test	189,708	229,084
合计	795,397	1,047,969

附加评测资源：

资源	规模	用途
模型选择集	5,000 行	用于选择 checkpoint；覆盖验证集 66 个来源身份及全部 38 项能力，每个母本组 1 行
校准集	2,000 行	预留用于概率校准（temperature / 阈值）；与模型选择集和测试集母本组不重叠

任务体系

每行归属一个决策桶。损失份额为各桶在总训练损失中的推荐占比，由 training_config.json 中的桶乘子实现。

桶	说明	训练行数	损失份额
A	搜索动作：下一步动作、停止/继续搜索、边际增益、下一跳查询选择	120,025	40%
B1	查询–文档质量：时效性、权威性、用户满意度	61,221	10%
B2	查询–文档相关性与排序	55,376	12%
C	证据与事实核查	41,288	12%
D1	句对与改写等价	65,465	8%
D2	查询评估：检索需求、信息量、噪声	42,729	8%
D3	查询意图分类	29,737	4%
E	通用理解：情感、自然语言推理、主题、偏好	51,527	6%

38 项能力的完整列表见英文部分。

搜索决策合成场景

为覆盖公开数据集中少有标注的决策环节，数据集包含 27,183 行专门构造的合成场景（input_origin = synth_v1），以及 10,109 行多维质量数据（synth_tas）。

轨道	场景	train	val	test
T1	状态条件化动作：证据充分性、冲突、预算与下一步行动	2,710	165	185
T2	SERP 选读：选择值得打开的搜索结果	494	28	32
T3	边际增益：新增信息、重复证据、冲突与时效变化	1,802	117	124
T4	下一跳查询：中文候选查询选择	3,663	440	366
T5	来源与时间：权威性与时效性判断	3,762	265	293
T6	失败恢复：搜索或读页失败、结果不可用、注入内容	5,297	313	366
T7	上下文改写：多轮对话中的意图保持	1,320	317	271
T10	停止边界：充分、过时、冲突与冗余证据	378	28	20
T11	澄清决策：何时向用户澄清及下一步行动	3,744	374	309

详见 SYNTHETIC_SCENARIOS.md。

数据来源

类别	代表性来源
检索与相关性	T2Ranking、DuRetrieval、Multi-CPR（电商/医疗/视频）、QBQTC、KUAKE-QQR、LCQMC、BQ Corpus、ATEC、AFQMC、CHIP-STS、PAWS-X（中文）、QReCC、对话改写等价、查询–文档质量标注
证据与核查	VitaminC、CFEVER、TrendFact、RAGTruth
搜索决策	搜索智能体轨迹、KUAKE-QIC、Search Arena、AgentRewardBench
搜索难负例	面向增益、下一跳、查询匹配与充分性的挖掘难负例，含中文翻译版与高难度子集
查询理解	CLINC150、Banking77、MASSIVE（中文）、查询信息量/检索需求/噪声评估
通用理解	CLUE（CMNLI、OCNLI、TNEWS、CSL、WSC、C3）、LogiQA（中文）、ChnSentiCorp、Weibo Senti 100k、Online Shopping 10 Cats、THUCNews 标题、Amazon Reviews Multi（中文）、豆瓣与京东评论、多轮情感对话、COIG-P、CValues、HelpSteer2 / HelpSteer3
合成数据	搜索决策场景 T1–T11 及多维质量数据（见上）

部分来源（如 Search Arena、AgentRewardBench、RAGTruth、QReCC 测试集）仅出现在 test 中，作为留出评测使用。

数据格式

{train,val,test}.jsonl 每行一个样本，包含 state、questions（问题字典）、answers（参考答案）和 difficulty（难度等级），示例见英文部分。

题型

类型	含义	criteria	答案
noul	命题是否为真	可选 {"true": ..., "false": ...}	noul：为真的概率
choice	从固定选项中单选	{选项: 描述}	choice、confidence、全部选项上的 probabilities
score	2–10 档有序打分	档位描述列表（低 → 高）	score（期望档位）、confidence、按档位的 probabilities

state 可以是字符串、JSON 对象或列表。一行可包含针对同一状态的多道问题，每道题独立作答。

索引文件。 {split}_index.jsonl 与 {split}.jsonl 逐行对齐，保存元数据。主要字段：

字段	说明
sample_id、source、input_origin	样本标识与来源
capability、task_family、decision_bucket	任务分类
training_lane、supervision、weak、label_tier、target_provenance	标签来源与监督强度
loss_weight、row_loss_scale	题级损失权重与行级缩放
evaluation_eligible、eval_track	各题是否参与评测及所属评测轨道
gold、confidence、teacher_reference_probability	参考答案与辅助教师概率
canonical_group、group_keys	防泄漏切分所用的分组键
language_heuristic、difficulty、token_preflight	语言、难度与长度统计

文件说明

文件	说明
train.jsonl、val.jsonl、test.jsonl	主数据
train_index.jsonl、val_index.jsonl、test_index.jsonl	逐行对齐的元数据
model_selection.jsonl、model_selection_index.jsonl	固定的 checkpoint 选择集
calibration_blind.jsonl、calibration_reference.jsonl	校准集：不含答案的题目及对应参考标签
training_config.json	参考训练配置（Qwen3-1.7B、LoRA、桶乘子）
budget.json	各桶损失预算
manifest.json	发布元数据与 SHA-256 校验值
data_report.json、DATA_REPORT.md	统计数据与详细构建报告
SYNTHETIC_SCENARIOS.md	合成场景说明
VERSION_HISTORY.md、VERSION_HISTORY.json	版本记录
changes.jsonl	相对上一版本的逐行变更记录

使用方法

读取方式见英文部分的 Python 示例。训练时建议每次前向只放一道题（独立上下文），每题损失权重为 loss_weight × row_loss_scale × 桶乘子。SearchJev 代码仓库提供了可直接读取本目录的参考训练与评测流程。

构建与质量控制

• 统一格式。 所有来源映射为「状态 / 类型化问题 / 答案」格式，并对题目指令做多样化改写，避免模板过拟合。
• 标签清洗。 疑似错误或存在争议的标签、未经判定的负例、口径未明的评分映射以及存在捷径特征（如输入中可见未来动作）的题目一律移出监督，不依据模型投票自动改标。
• 防泄漏。 基于母本、翻译与挖掘父本及相同输入做分组切分，并去除完全重复的输入。
• 长度控制。 每道保留题均以完整状态做 6,144 token 预检，超长题整题移除而非截断证据。
• 损失预算。 按桶计算乘子，使实际损失份额与目标 A/B1/B2/C/D1/D2/D3/E = 40/10/12/12/8/8/4/6% 一致。

局限性

• 训练目标为源自原始标注、示范行为或构造规则的硬参考标签。答案中的 confidence = 1.0 仅表示目标编码，不代表实际准确率；校准概率应在校准集上拟合。
• 动作模仿类题目的标签表示示范者的实际行为，不一定是最优策略。
• A 桶仍主要来自派生的英文多跳数据，训练集中 A 桶的中文行约占 19%。
• 模型选择集为分层抽样，不能作为自然分布下准确率的估计。

许可

本数据集由多个公开数据集及合成数据加工而成，各子集仍受其原始来源的许可与使用条款约束，使用前请查阅相应上游许可。本数据集仅供研究使用。

引用

如使用本数据集，请引用：

@misc{cao2026searchjev,
  title  = {{SearchJev}: A Fast and Calibrated System-1 Model for Search Agents},
  author = {Cao, Congfeng and Zuo, Lipeng and Papakostas, Konstantinos and Xu, Qiwei and Xu, Songwei and
            Zhou, Lun and Ren, Zhaochun and Lyu, Yougang and Yan, Xiaohui},
  year   = {2026}
}