Unverified50% confidenceBenchmarkExact time
在评估阶段对候选关系标签进行随机打乱后，该模型的F1从37%骤降到14%
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/5/2026
First Seen
Valid until: 11/3/2026
Sources
Related Claims
Unverified对于参数量大于30B的模型，4-bit量化造成的性能损失通常在基准测试中下降1-3%，但对小模型影响更显著59% similarUnverified当相关文档被放置在上下文中间位置时，模型正确回答率相比放在首尾位置下降高达20-30个百分点59% similarUnverifiedH-1B普通类别的抽签录取率近年来已降至约15%—20%59% similarUnverified量化推理从FP16降至FP8可将显存需求减半，从FP8降至FP4可进一步减半59% similarUnverified轻量换来的代价是耐用性和抗穿刺性下降，口碑中扎胎和慢撒气的反馈明显多于标准版SV1559% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/692516API
curl https://kongchang.com/api/v1/knowledge/claims/692516MCP
get_claim(id=692516)