原始 2,000 张图库：微调对照实验

正式实验 8 组，检查点 24 个。训练及验证合计 26.23 小时；表中试验耗时会按检查点重复展示，不应重复相加。

所有模型由验证集选定。下表分数为 0–1；F1 阈值只在验证集确定。仅衡量类别级相关性，不证明颜色、材质、意图识别或实例检索质量。

测试指标 CSV · 全部超参 CSV · 类别粗筛 CSV · 逐查询变化 CSV · 完整 JSON · Top-20 与错误明细清单

固定完整测试图库

variant	mode	primary_five_map	primary_five_p1	primary_five_p5	primary_five_calibrated_f1	all_class_macro_map
embedding-base	text	0.656300	1.000000	0.840000	0.616808	0.873258
embedding-base	image	0.382749	0.665000	0.604000	0.318470	0.714324
embedding-best	text	0.935520	1.000000	1.000000	0.888363	0.975528
embedding-best	image	0.855698	0.845000	0.862000	0.834013	0.939170
qwen-base	text	0.472286	0.666667	0.840000	0.606169	0.658120
qwen-base	image	0.268696	0.455000	0.443000	0.356405	0.479526
qwen-old	text	0.412443	0.933333	0.906667	0.515864	0.610676
qwen-old	image	0.202044	0.365000	0.372000	0.289197	0.417355
qwen-best	text	0.806077	0.800000	0.813333	0.867595	0.913232
qwen-best	image	0.773066	0.825000	0.823000	0.818583	0.881038

粗筛与输出格式

variant	primary_five_macro_f1	macro_f1	valid_json_rate
qwen-base	0.627312	0.775610	0.846000
qwen-old	0.572735	0.738933	0.801500
qwen-best	0.873757	0.943040	1.000000

超参数及轮数

两类模型的 validation_score 定义不同，不横向比较该列。

family	learning_rate	temperature	mode	epoch_requested	actual_epoch	validation_score	selected
embedding	0.000001	0.050000	full-text-and-vision	0.500000	0.501333	0.828652	False
embedding	0.000001	0.050000	full-text-and-vision	1.000000	1.001778	0.918740	False
embedding	0.000001	0.050000	full-text-and-vision	2.000000	2.000000	0.944694	False
embedding	0.000002	0.050000	full-text-and-vision	0.500000	0.501333	0.921111	False
embedding	0.000002	0.050000	full-text-and-vision	1.000000	1.001778	0.947410	False
embedding	0.000002	0.050000	full-text-and-vision	2.000000	2.000000	0.950651	False
embedding	0.000005	0.050000	full-text-and-vision	0.500000	0.501333	0.936185	False
embedding	0.000005	0.050000	full-text-and-vision	1.000000	1.001778	0.939843	False
embedding	0.000005	0.050000	full-text-and-vision	2.000000	2.000000	0.945917	False
embedding	0.000002	0.100000	full-text-and-vision	0.500000	0.501333	0.934599	False
embedding	0.000002	0.100000	full-text-and-vision	1.000000	1.001778	0.953000	True
embedding	0.000002	0.100000	full-text-and-vision	2.000000	2.000000	0.950561	False
qwen	0.000005		llm-lora	0.500000	0.501333	0.870483	False
qwen	0.000005		llm-lora	1.000000	1.002667	0.911801	False
qwen	0.000005		llm-lora	2.000000	2.000000	0.914182	False
qwen	0.000010		llm-lora	0.500000	0.501333	0.868657	False
qwen	0.000010		llm-lora	1.000000	1.002667	0.921549	True
qwen	0.000010		llm-lora	2.000000	2.000000	0.911457	False
qwen	0.000020		llm-lora	0.500000	0.501333	0.887372	False
qwen	0.000020		llm-lora	1.000000	1.002667	0.917913	False
qwen	0.000020		llm-lora	2.000000	2.000000	0.907592	False
qwen	0.000010		mlp-llm	0.500000	0.501333	0.877523	False
qwen	0.000010		mlp-llm	1.000000	1.002667	0.906233	False
qwen	0.000010		mlp-llm	2.000000	2.000000	0.908266	False

分组与旧训练重叠敏感性分析

fine 是小目标分组。排除已检出的旧训练潜在重叠不等于证明模型从未见过图片。

variant	scope	mode	gallery_size	primary_five_map	primary_five_calibrated_f1
embedding-base	normal	text	1000	0.703885	0.641656
embedding-base	normal	image	1000	0.446146	0.375594
embedding-base	fine	text	1000	0.634467	0.591956
embedding-base	fine	image	1000	0.359361	0.278950
embedding-base	overlap-excluded	text	1851	0.645092	0.610049
embedding-base	overlap-excluded	image	1851	0.390766	0.322467
embedding-best	normal	text	1000	0.948710	0.895679
embedding-best	normal	image	1000	0.879081	0.836369
embedding-best	fine	text	1000	0.927385	0.877996
embedding-best	fine	image	1000	0.832085	0.821874
embedding-best	overlap-excluded	text	1851	0.934617	0.878466
embedding-best	overlap-excluded	image	1851	0.852805	0.833660
qwen-base	normal	text	1000	0.546806	0.680162
qwen-base	normal	image	1000	0.369242	0.477627
qwen-base	fine	text	1000	0.406492	0.516782
qwen-base	fine	image	1000	0.183108	0.245069
qwen-base	overlap-excluded	text	1851	0.470990	0.601156
qwen-base	overlap-excluded	image	1851	0.277999	0.365364
qwen-old	normal	text	1000	0.502643	0.586037
qwen-old	normal	image	1000	0.308166	0.416859
qwen-old	fine	text	1000	0.323778	0.418089
qwen-old	fine	image	1000	0.128480	0.183425
qwen-old	overlap-excluded	text	1851	0.425897	0.528983
qwen-old	overlap-excluded	image	1851	0.226758	0.314223
qwen-best	normal	text	1000	0.814236	0.893650
qwen-best	normal	image	1000	0.806960	0.850612
qwen-best	fine	text	1000	0.822566	0.839208
qwen-best	fine	image	1000	0.740729	0.785257
qwen-best	overlap-excluded	text	1851	0.791875	0.854179
qwen-best	overlap-excluded	image	1851	0.766815	0.811997

逐查询增益与退化

完整逐查询差值见 paired_changes.csv；负值照常保留。只有单次种子实验，不据此声称统计显著。

embedding-base -> embedding-best

最大退化与最大提升各 5 个查询：

mode	query_id	label	before_ap	after_ap	ap_delta
image	n07747607_19866	orange	0.772796	0.333658	-0.439139
image	n04225987_2297	skateboard	0.446431	0.097142	-0.349289
image	n07747607_2730	orange	0.895431	0.565812	-0.329619
image	n07747607_19598	orange	0.653102	0.323506	-0.329596
image	n04251791_8522	snowboard	0.451578	0.140805	-0.310773
image	n03147509_5048	cup	0.092150	0.948425	0.856275
image	n03147509_17845	cup	0.101657	0.959643	0.857985
image	n03001627_9421	chair	0.093315	0.958509	0.865194
image	n03147509_5550	cup	0.041811	0.949365	0.907554
image	n03001627_5048	chair	0.038084	0.962623	0.924540

qwen-base -> qwen-best

最大退化与最大提升各 5 个查询：

mode	query_id	label	before_ap	after_ap	ap_delta
image	n02121620_29395	cat	0.908015	0.000000	-0.908015
image	n02881193_8326	bowl	0.510895	0.039889	-0.471006
image	n02881193_3035	bowl	0.439396	0.024893	-0.414504
image	n03761084_10934	microwave	0.080154	0.024781	-0.055373
image	n03761084_6238	microwave	0.084259	0.029762	-0.054497
image	n02834778_12089	bicycle	0.000000	1.000000	1.000000
image	n02834778_5634	bicycle	0.000000	1.000000	1.000000
image	n02834778_1832	bicycle	0.000000	1.000000	1.000000
image	n02834778_5383	bicycle	0.000000	1.000000	1.000000
image	n02834778_98	bicycle	0.000000	1.000000	1.000000

qwen-old -> qwen-best

最大退化与最大提升各 5 个查询：

mode	query_id	label	before_ap	after_ap	ap_delta
image	n02121620_29395	cat	0.871265	0.000000	-0.871265
image	n02881193_8326	bowl	0.534850	0.039889	-0.494962
image	n03862676_3543	oven	0.134556	0.032630	-0.101926
image	n03761084_15107	microwave	0.076969	0.016305	-0.060664
image	n03761084_6238	microwave	0.089617	0.029762	-0.059855
image	n02834778_5634	bicycle	0.000000	1.000000	1.000000
image	n02834778_1832	bicycle	0.000000	1.000000	1.000000
image	n02834778_5383	bicycle	0.000000	1.000000	1.000000
image	n02834778_98	bicycle	0.000000	1.000000	1.000000
image	n02834778_10841	bicycle	0.000000	1.000000	1.000000

基准采用固定 15 类映射粗筛及纯参考图查询；网页的开放类别语义筛选、意图识别和图文融合应另行验收。新指标不冒充旧脚本在测试集选阈值得到的历史数字。

实验图表

下载 PDF

下载 PDF

下载 PDF