Appear2Meaning Human Annotation

Files and assignments / 文件与分工

Each Excel workbook contains 75 objects: five objects from each of the 15 region × object type strata. Each row includes an embedded primary/front-view image immediately before the reference title. The two copies for a group contain the same objects, references, and model outputs. Only the annotator column differs.

每份 Excel 工作簿包含 75 件藏品，即 15 个 region × object type 类别各 5 件。每行在 reference title 前嵌入一张藏品的主要／正面图片。同一组的两份文件包含完全相同的藏品、参考标签和模型输出，仅末列 annotator 不同。

Group	Models	Annotators
Group 1	Qwen-VL-Max; Qwen3-VL-8B; GPT-4.1-mini	A, B
Group 2	Qwen3-VL-Plus; Qwen3-VL-32B; GPT-5.4-mini	B, C
Group 3	Qwen3-VL-Flash; Pixtral-12B; Claude Haiku 4.5	A, C

Annotation guideline (English)

1. Annotate the five fields independently: title, culture, period, origin, and creator.
2. For each model and field, enter exactly 1 or 0 in the adjacent _match_0_or_1 column.
   • 1: the model output is fully or broadly consistent with at least one reference label in meaning and scope.
   • 0: the model output is incorrect, contradictory, unsupported, only says “unknown/not sure” when the reference contains an answer, or does not meaningfully overlap with the reference.
3. Exact wording is not required. Accept ordinary aliases, spelling or transliteration variants, historical and modern place names, adjectival and noun forms, and equivalent chronological expressions when they refer to the same information.
4. Reference alternatives are separated by ||. A model output may receive 1 when it agrees with any applicable reference alternative. It does not need to reproduce every alternative.
5. Evaluate only the current field. Correct information placed in the wrong field does not make that field correct.
6. For period, accept a compatible century, dynasty, reign, date range, or historical period when its range is broadly consistent with the reference. Mark a conflicting or materially different period as 0.
7. For origin, a compatible city, region, country, or historical/modern place name may receive 1 when it identifies the same origin without contradiction.
8. For creator, accept established aliases and the correctly identified person, workshop, factory, manufactory, kiln, studio, or attributed group. “Unknown” receives 1 only when the reference explicitly indicates an unknown or anonymous creator.
9. If the reference cell is empty, leave the corresponding score cell empty. Do not infer a reference from the sampling region, object type, description, or another field.
10. Annotate independently. Do not consult the other annotator’s decisions before both copies are complete.

Examples

• Reference culture Japanese; output Japan → 1.
• Reference period Ming dynasty || 14th–17th century; output Ming period → 1.
• Reference origin Strasbourg || Alsace || France; output Alsace, France → 1.
• Reference creator Ding kiln; output Unknown → 0.
• Reference title Butter Pat; output describing an unrelated vase → 0.

标注规范（中文）

1. 分别独立判断五个字段：title、culture、period、origin 和 creator。
2. 对每个模型的每个字段，只在相邻的 _match_0_or_1 列中填写 1 或 0。
   • 1：模型输出在含义和范围上完全正确，或与至少一个 reference label 基本吻合。
   • 0：模型输出错误、相互矛盾、缺乏依据、与 reference 没有实质重合，或者在 reference 已有答案时仅回答 “unknown/not sure”。
3. 不要求文字完全相同。若普通别名、拼写或转写变体、古今地名、名词与形容词形式，以及不同年代表示法指向相同信息，应判为 1。
4. 同一 reference 单元格中的多个标签以 || 分隔。模型输出只要与其中一个适用标签吻合即可判为 1，不要求覆盖全部标签。
5. 只判断当前字段。即使某项信息本身正确，但被模型填写在错误字段中，也不能据此将当前字段判为正确。
6. 对 period，若世纪、朝代、年号、年代范围或历史时期与 reference 的时间范围基本一致，可判为 1；若年代冲突或存在实质差异，则判为 0。
7. 对 origin，若城市、地区、国家或古今地名指向相同来源且不存在矛盾，可判为 1。
8. 对 creator，可接受创作者的通用别名，以及正确的个人、工坊、工厂、制造厂、窑口、工作室或归属群体。“Unknown” 只有在 reference 明确标注创作者未知或匿名时才判为 1。
9. 如果 reference 单元格为空，请将对应评分单元格留空。不得根据抽样使用的 region、object type、description 或其他字段自行补充 reference。
10. 两位标注者应独立完成各自文件。在双方完成之前，不应查看或讨论另一位标注者的判断。

示例

• Reference culture 为 Japanese，模型输出 Japan → 1。
• Reference period 为 Ming dynasty || 14th–17th century，模型输出 Ming period → 1。
• Reference origin 为 Strasbourg || Alsace || France，模型输出 Alsace, France → 1。
• Reference creator 为 Ding kiln，模型输出 Unknown → 0。
• Reference title 为 Butter Pat，模型却描述为无关的花瓶 → 0。

Column layout / 列结构

The first column is the museum description, the second contains the embedded object image, and the third is the reference title. For Title, columns 4, 6, and 8 contain the three model outputs; columns 5, 7, and 9 are their annotation cells. Culture, Period, Origin, and Creator follow the same reference/output/score pattern. Tracking columns and the assigned annotator appear at the end.

第一列为博物馆描述，第二列为嵌入的藏品图片，第三列为 reference title。Title 部分的第 4、6、8 列依次为三个模型的输出，第 5、7、9 列为对应的人工评分。Culture、Period、Origin 和 Creator 使用相同的 reference／模型输出／评分结构。文件末尾保留对象追踪信息和指定 annotator。

LLM-as-a-Judge evaluation / LLM Judge评估

The evaluation contains 675 records: 225 sampled objects and three evaluated model outputs per object. Each judge independently scores Title, Culture, Period, Origin, and Creator as 0 or 1. Valid is the number of outputs that contain a parseable five-field response. Invalid generation cases are excluded from every accuracy denominator. A structurally valid response consisting of unknown or not sure remains valid. When a reference field is empty, an explicitly unknown or substantively uncertain prediction receives 1, while a definite prediction receives 0.

本次评估包含675条记录，即225件抽样藏品，每件对应三个待评估模型输出。四个Judge分别独立判断Title、Culture、Period、Origin和Creator，结果为0或1。Valid表示能够解析为五字段结构化回复的输出数量；生成阶段的无效case不进入任何正确率的分母。格式有效但全部回答unknown或not sure的输出仍属于有效case。当某个reference字段为空时，明确回答未知或使用实质性不确定描述的预测记为1，给出确定答案则记为0。

Files / 文件：

• sampled_675_records.json: sampled records, descriptions, reference labels, and evaluated model outputs.
• run_three_judges.py and judge_prompt.md: evaluation code and prompt. API keys are not included.
• open_source.env.example: configuration example for Llama and Qwen served through OpenAI-compatible endpoints such as vLLM, SGLang, or a compatible local/remote gateway. Set LLAMA_BASE_URL, LLAMA_MODEL, QWEN_BASE_URL, and QWEN_MODEL; API keys are optional for unauthenticated local endpoints. Run --judges llama_open_source qwen_open_source to evaluate with both open-source interfaces. The default command continues to run only the original three cloud Judges.
• results/: the four independent Judge result files.
• judge_summary.csv: unrounded summary values for the original three Judge tables below. The Qwen table is calculated directly from its released result file using the same valid-output rule.

(a) GPT-4.1-mini as Judge

Model	Valid	Title	Culture	Period	Origin	Creator
Qwen-VL-Max	74	0.284	0.486	0.541	0.392	0.257
Qwen3-VL-Plus	69	0.275	0.522	0.478	0.449	0.217
Qwen3-VL-Flash	72	0.347	0.597	0.569	0.528	0.403
Qwen3-VL-8B	75	0.160	0.440	0.413	0.333	0.293
Qwen3-VL-32B	75	0.373	0.547	0.493	0.467	0.307
Pixtral-12B	67	0.284	0.463	0.478	0.418	0.448
GPT-4.1-mini	75	0.333	0.493	0.547	0.467	0.320
GPT-5.4-mini	75	0.320	0.573	0.547	0.507	0.293
Claude Haiku 4.5	74	0.270	0.405	0.432	0.311	0.419

(b) DeepSeek-V4-Flash as Judge

The API accepts deepseek-v4-flash as an alias and reports the resolved model as deepseek-flash in the returned records.

Model	Valid	Title	Culture	Period	Origin	Creator
Qwen-VL-Max	74	0.297	0.514	0.500	0.432	0.432
Qwen3-VL-Plus	69	0.406	0.522	0.377	0.449	0.246
Qwen3-VL-Flash	72	0.319	0.611	0.542	0.528	0.500
Qwen3-VL-8B	75	0.213	0.440	0.347	0.347	0.400
Qwen3-VL-32B	75	0.333	0.480	0.453	0.507	0.307
Pixtral-12B	67	0.328	0.478	0.493	0.478	0.478
GPT-4.1-mini	75	0.387	0.480	0.480	0.467	0.373
GPT-5.4-mini	75	0.360	0.560	0.467	0.507	0.347
Claude Haiku 4.5	74	0.324	0.500	0.500	0.405	0.446

(c) Gemini-3.5-Flash-Lite as Judge

Model	Valid	Title	Culture	Period	Origin	Creator
Qwen-VL-Max	74	0.216	0.486	0.595	0.527	0.338
Qwen3-VL-Plus	69	0.333	0.536	0.478	0.580	0.174
Qwen3-VL-Flash	72	0.347	0.597	0.597	0.681	0.458
Qwen3-VL-8B	75	0.213	0.427	0.453	0.453	0.280
Qwen3-VL-32B	75	0.293	0.480	0.467	0.547	0.280
Pixtral-12B	67	0.284	0.493	0.522	0.612	0.373
GPT-4.1-mini	75	0.320	0.480	0.600	0.587	0.347
GPT-5.4-mini	75	0.320	0.573	0.560	0.613	0.173
Claude Haiku 4.5	74	0.311	0.459	0.500	0.527	0.351

(d) Qwen3-30B-A3B-Instruct-2507 as Judge

This Judge was run locally from the released open-source model.

Model	Valid	Title	Culture	Period	Origin	Creator
Qwen-VL-Max	74	0.162	0.486	0.554	0.378	0.514
Qwen3-VL-Plus	69	0.188	0.507	0.333	0.362	0.217
Qwen3-VL-Flash	72	0.306	0.542	0.542	0.542	0.444
Qwen3-VL-8B	75	0.173	0.467	0.480	0.413	0.427
Qwen3-VL-32B	75	0.320	0.600	0.547	0.533	0.400
Pixtral-12B	67	0.164	0.478	0.433	0.418	0.418
GPT-4.1-mini	75	0.187	0.480	0.533	0.480	0.360
GPT-5.4-mini	75	0.240	0.560	0.467	0.507	0.293
Claude Haiku 4.5	74	0.230	0.419	0.419	0.392	0.392

Agreement among the four LLM Judges / 四个 LLM Judge 的一致性

Agreement is calculated over the 656 valid model outputs; the 19 invalid generation cases are excluded. Four-way agreement is the proportion of cases for which all four Judges assign the same binary score. Pairwise columns report raw agreement. Fleiss' κ measures four-Judge agreement after accounting for agreement expected by chance.

一致性基于 656 条有效模型输出计算，排除 19 条生成阶段的无效 case。四方完全一致率表示四个 Judge 给出相同二元评分的比例；两两一致率为原始一致率；Fleiss' κ 则进一步扣除了随机一致的影响。

Field	Four-way	Mean pairwise	Fleiss' κ	GPT–DeepSeek	GPT–Gemini	GPT–Qwen	DeepSeek–Gemini	DeepSeek–Qwen	Gemini–Qwen
Title	0.768	0.872	0.684	0.895	0.877	0.873	0.912	0.826	0.848
Culture	0.931	0.963	0.925	0.966	0.973	0.962	0.976	0.944	0.956
Period	0.782	0.880	0.760	0.886	0.884	0.877	0.889	0.866	0.880
Origin	0.774	0.878	0.755	0.909	0.846	0.896	0.870	0.899	0.846
Creator	0.738	0.857	0.687	0.890	0.858	0.863	0.843	0.878	0.809

Culture remains the most stable field: all four Judges agree on 93.1% of valid outputs and Fleiss' κ is 0.925. Period and Origin also retain relatively strong agreement. The additional Qwen Judge reduces agreement most clearly for Title, where four-way agreement is 76.8% and κ is 0.684. Creator remains the least stable field, with 73.8% four-way agreement and κ = 0.687. Overall, the four Judges are highly consistent on Culture and reasonably consistent on the remaining fields, but Title and Creator require greater caution and human validation.

Culture 仍是最稳定的字段：四个 Judge 在 93.1% 的有效输出上给出相同判断，Fleiss' κ 为 0.925。Period 和 Origin 也保持了较高一致性。加入 Qwen Judge 后，Title 的一致性下降最明显，四方完全一致率为 76.8%，κ 为 0.684。Creator 仍是最不稳定的字段，四方完全一致率为 73.8%，κ 为 0.687。总体而言，四个 Judge 对 Culture 的判断高度一致，对其余字段也有较好一致性，但 Title 和 Creator 更需要谨慎解释并通过人工评估验证。