垂直专项评测
模态: 文本 (Text)
更新频率: weekly
● 自动定时拉取
Vectara Hallucination Leaderboard
专注于大语言模型摘要生成场景下事实一致性与幻觉率的量化评测。
💡 评测口径提示： 幻觉率越低越优，直连 Vectara 官方开源基准库提取。
最后核验：2026-10-26
模型最新排名快照
共收录 15 款模型。数据均保留原榜单给出的分数与分项维度。
| 名次 | 模型名称 | 统一标识 | 机构 | 核心分值 | 分项指标明细 | 备注说明 | 跨榜对比 |
|---|---|---|---|---|---|---|---|
| 1 | antgroup/finix_s1_32b | antgroup-finix-s1-32b | antgroup | 98.2 | Hallucination Rate: 1.8% Factual Consistency Rate: 98.2% Answer Rate: 99.5% Avg Summary Length: 172.4 | 事实一致性: 98.2%, 幻觉率: 1.8%, 回答率: 99.5% | 对比 → |
| 2 | openai/gpt-5.4-nano-2026-03-17 | openai-gpt-5.4-nano-2026-03-17 | openai | 96.9 | Hallucination Rate: 3.1% Factual Consistency Rate: 96.9% Answer Rate: 100% Avg Summary Length: 144.4 | 事实一致性: 96.9%, 幻觉率: 3.1%, 回答率: 100% | 对比 → |
| 3 | google/gemini-2.5-flash-lite | google-gemini-2.5-flash-lite | 96.7 | Hallucination Rate: 3.3% Factual Consistency Rate: 96.7% Answer Rate: 99.5% Avg Summary Length: 95.7 | 事实一致性: 96.7%, 幻觉率: 3.3%, 回答率: 99.5% | 对比 → | |
| 4 | microsoft/Phi-4 | microsoft-phi-4 | microsoft | 96.3 | Hallucination Rate: 3.7% Factual Consistency Rate: 96.3% Answer Rate: 80.7% Avg Summary Length: 120.9 | 事实一致性: 96.3%, 幻觉率: 3.7%, 回答率: 80.7% | 对比 → |
| 5 | meta-llama/Llama-3.3-70B-Instruct-Turbo | meta-llama-llama-3.3-70b-instruct-turbo | meta-llama | 95.9 | Hallucination Rate: 4.1% Factual Consistency Rate: 95.9% Answer Rate: 99.5% Avg Summary Length: 64.6 | 事实一致性: 95.9%, 幻觉率: 4.1%, 回答率: 99.5% | 对比 → |
| 6 | deepseek-ai/DeepSeek-V4.1-Flash | deepseek-ai-deepseek-v4.1-flash | deepseek-ai | 95.8 | Hallucination Rate: 4.2% Factual Consistency Rate: 95.8% Answer Rate: 99.9% Avg Summary Length: 108.2 | 事实一致性: 95.8%, 幻觉率: 4.2%, 回答率: 99.9% | 对比 → |
| 7 | snowflake/snowflake-arctic-instruct | snowflake-snowflake-arctic-instruct | snowflake | 95.7 | Hallucination Rate: 4.3% Factual Consistency Rate: 95.7% Answer Rate: 62.7% Avg Summary Length: 81.4 | 事实一致性: 95.7%, 幻觉率: 4.3%, 回答率: 62.7% | 对比 → |
| 8 | google/gemma-3-12b-it | google-gemma-3-12b-it | 95.6 | Hallucination Rate: 4.4% Factual Consistency Rate: 95.6% Answer Rate: 97.4% Avg Summary Length: 89.7 | 事实一致性: 95.6%, 幻觉率: 4.4%, 回答率: 97.4% | 对比 → | |
| 9 | mistralai/mistral-large-2411 | mistralai-mistral-large-2411 | mistralai | 95.5 | Hallucination Rate: 4.5% Factual Consistency Rate: 95.5% Answer Rate: 99.9% Avg Summary Length: 85.0 | 事实一致性: 95.5%, 幻觉率: 4.5%, 回答率: 99.9% | 对比 → |
| 10 | qwen/qwen3-8b | qwen-qwen3-8b | qwen | 95.2 | Hallucination Rate: 4.8% Factual Consistency Rate: 95.2% Answer Rate: 99.9% Avg Summary Length: 83.6 | 事实一致性: 95.2%, 幻觉率: 4.8%, 回答率: 99.9% | 对比 → |
| 11 | amazon/nova-pro-v1:0 | amazon-nova-pro-v1-0 | amazon | 94.9 | Hallucination Rate: 5.1% Factual Consistency Rate: 94.9% Answer Rate: 99.3% Avg Summary Length: 66.2 | 事实一致性: 94.9%, 幻觉率: 5.1%, 回答率: 99.3% | 对比 → |
| 12 | amazon/nova-2-lite-v1:0 | amazon-nova-2-lite-v1-0 | amazon | 94.9 | Hallucination Rate: 5.1% Factual Consistency Rate: 94.9% Answer Rate: 99.6% Avg Summary Length: 94.1 | 事实一致性: 94.9%, 幻觉率: 5.1%, 回答率: 99.6% | 对比 → |
| 13 | mistralai/mistral-small-2501 | mistralai-mistral-small-2501 | mistralai | 94.9 | Hallucination Rate: 5.1% Factual Consistency Rate: 94.9% Answer Rate: 97.9% Avg Summary Length: 98.8 | 事实一致性: 94.9%, 幻觉率: 5.1%, 回答率: 97.9% | 对比 → |
| 14 | ibm-granite/granite-4.0-h-small | ibm-granite-granite-4.0-h-small | ibm-granite | 94.8 | Hallucination Rate: 5.2% Factual Consistency Rate: 94.8% Answer Rate: 100% Avg Summary Length: 107.4 | 事实一致性: 94.8%, 幻觉率: 5.2%, 回答率: 100% | 对比 → |
| 15 | google/gemma-4-26b-a4b-it | google-gemma-4-26b-a4b-it | 94.8 | Hallucination Rate: 5.2% Factual Consistency Rate: 94.8% Answer Rate: 99.8% Avg Summary Length: 67.1 | 事实一致性: 94.8%, 幻觉率: 5.2%, 回答率: 99.8% | 对比 → |
快照采集说明： 直接实时抓取自 Vectara 官方 HHEM 幻觉评测基准库（涵盖最新主流大语言模型摘要一致性与幻觉率测试）。
📊 本榜单指标与评价方法说明
选型决策建议：结合具体应用场景（如代码辅助、企业知识库检索、对话推理）参考专有指标，切勿以单一数值判定模型整体优劣。