Primary sourceJerry Liu
View sourceEvent ·Multi-source
LlamaIndex 发布 ParseBench-2 与 OCRia-1.5 文档解析模型
LlamaIndex 推出文档解析基准 ParseBench-2,包含 4.3k+ 页文档,比 ParseBench-1 翻倍,并新增表单(forms)维度。基准为 330+ 种错误类型逐一标注严重程度分数,再据此调校各项指标,使其与人工判断高度一致。 现有基准的评测方式各有短板:OmniDocBench 依赖脆弱的确定性指标,ParseArena 采用 LLM-as-a-judge,结果粗糙、随机且受提示词和模型偏见影响。ParseBench-1 已积累 50k+ 月下载量,评测过 150+ 个 OCR 方案,被 NVIDIA、Cohere、Perplexity、Aleph Alpha 等公司采用。 同期发布 OCRia-1.5 文档解析模型家族,驱动 LlamaParse 全部解析模式,在复杂表格、表单、图表、版式和边界框标注上达到当前最优水平。技术报告、博客和官网均已公开。
Current read
做 RAG 或文档处理的开发者该看看。旧基准要么指标太脆要么靠 LLM 打分不靠谱,这次按 330 多种错误类型标了严重度来调指标,评测可信多了。附带的 OCRia-1.5 处理复杂表格和表单更强,可以直接上 LlamaParse 试。
9 sources55° AI heatLast updated 2026/11/12 16:46:42
Evidence
9 sourcesRelated source 1NVIDIA AI
View sourceRelated source 2Decoder
View sourceRelated source 3rohanpaul_ai
View sourceRelated source 4vLLM
View sourceRelated source 5Jim Fan
View sourceRelated source 6AI21 Labs
View sourceRelated source 7orange.ai
View sourceRelated source 8Aravind Srinivas
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.