t-tech/SynthComp
收藏资源简介:
SynthComp是一个合成的类BrowseComp-Plus检索基准,旨在评估语言模型和搜索代理是否能够为困难的组合问题检索完整的支持证据集。该基准包含两个语言版本:SynthComp-Ru(俄语)和SynthComp-En(英语),每个版本包含395个多跳短答案问题。与手动编写的基准(如TRuST)不同,SynthComp是完全合成的:问题由LLM作曲器基于从真实网络语料库中采样的受控子图生成,并通过程序和模型门控的对抗级联进行过滤。目标是仅保留那些所有支持证据都是必要的、没有单个广泛查询能直接检索到答案、且在最终固定索引中不存在替代检索路径的问题。基准包含四种合成挑战机制:桥接(通过共享实体或桥接事实进行多跳链接)、星型(约束的交集)、容器(候选集构建)和时序(时间依赖问题)。数据集的构建过程包括子图挖掘、增量组合、转述事实图、可解性门控、表面泄漏门控、广泛搜索捷径门控、k-of-k消融门控、后修剪、去重和碰撞清理、干扰项重新挖掘、泄漏筛查、难度分层以及检索索引构建。评估使用原始的BrowseComp-Plus评估框架,要求解密基准文件并映射到BrowseComp-Plus仓库中运行。数据集预览包含加密的问题和答案,仅供本地检查使用,不应用于模型训练。
SynthComp is a synthetic BrowseComp-Plus-like retrieval benchmark designed to evaluate whether language models and search agents can retrieve a complete set of supporting evidence for hard compositional questions. The benchmark contains two language versions: SynthComp-Ru and SynthComp-En, each with 395 multi-hop short-answer questions. Unlike manually written benchmarks such as TRuST, SynthComp is fully synthetic. Questions are generated by an LLM composer over controlled subgraphs sampled from real web corpora, and then filtered through an adversarial cascade of programmatic and model-based gates. The goal is to keep only questions where all supporting evidence is necessary, no single broad query retrieves the answer directly, and no alternative retrievable path to the answer exists in the final fixed index. The benchmark includes 4 synthetic challenge mechanisms: Bridge (multi-hop linking through shared entities or bridge facts), Star (intersection of constraints), Container (candidate-set construction), and Temporal (time-dependent questions). The construction pipeline involves subgraph mining, incremental composition, paraphrased fact graph, solvability gate, surface-leak gate, broad-search shortcut gate, k-of-k ablation gate, post-pruning, deduplication and collision cleanup, distractor re-mining, leak screening, difficulty stratification, and retrieval index construction. Evaluation uses the original BrowseComp-Plus evaluation framework, requiring decryption of benchmark files and mapping into the BrowseComp-Plus repository. Dataset previews contain encrypted questions and answers for local inspection only and should not be used for model training.




