File size: 14,736 Bytes
907f90f 9afeefe 907f90f 9afeefe | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 | {
"name": "TableVQA-Bench",
"release_date": "2024-04-01",
"subsets": {
"vwtq": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "short_answer_qa",
"score_pipeline": [
"exact-match",
"rule-match"
],
"score_protocol": {
"reference": "Official naver-ai/tablevqabench src/evaluate_tablevqa_dataset.py:17-33 (per-split dispatch) — vwtq/vwtq_syn use WikiTableQuestions denotation matching (to_value_list/check_denotation with number/date normalization); vtabfact scores lowercase 'true'/'false' containment in the prediction (evaluate_tabfact, lines 45-64); fintabnetqa uses normalized exact match with $/unit-word stripping ('relieved accuracy'). Ported in VLMEvalKit@vlmeval/dataset/image_vqa.py:2366-2396 + vlmeval/dataset/utils/tablevqabench.py:54-158. Headline = mean accuracy per split.",
"note": "Fully deterministic (rule) but with per-split normalization the generic exact/template chain will NOT reproduce: WTQ denotation normalization and fintabnet unit stripping. vtabfact gt is stored as 'true'/'false' (mapped from source gt '1'/'0'), matching what the official evaluator expects from a generative model: evaluate_tabfact (src/evaluate_tablevqa_dataset.py:56-64) lowercases the prediction and scores \"'true' in pred and gt==1\" / \"'false' in pred and gt==0\" as correct, and the official VTABFACT_PROMPT (src/example_prompts.py:21-29) instructs 'You should only answer true or false.'. Official also uses split-specific prompts (VTABFACT_PROMPT etc., image_vqa.py:2398-2414) whereas the published prompt_template is the generic short-answer prompt. Official reporting is per-split accuracy averaged over 4 splits; published copy stores splits as HF splits under one subset."
},
"prompt_template": "<image>You are asked to answer questions asked on an image.\nYou should answer the question with a single word.\nExample: \nQuestion: what was the only year mr. wu competed in the olympic games?\nAnswer: 2004\nQuestion: which township in pope county, arkansas has the least amount of water area?\nAnswer: Freeman\nIf you have multiple answers, please separate them with || marks. Example: Apple||Banana||Tomato\n\nQuestion: {{ question }}\nAnswer:",
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "qa_id"
},
"question": {
"from": "question"
},
"answer": {
"from": "gt",
"optional": true,
"note": "vtabfact split only: source gt '1'/'0' mapped to 'true'/'false' per the official evaluator (naver-ai/tablevqabench src/evaluate_tablevqa_dataset.py:56-64: lowercase 'true'/'false' containment vs gt 1/0; src/example_prompts.py:21-29: 'You should only answer true or false.'). vwtq/vwtq_syn/fintabnetqa gts are verbatim source values."
},
"source": {
"format": "huggingface",
"url": {
"vwtq": "https://huggingface.co/datasets/terryoo/TableVQA-Bench"
}
}
},
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/naver-ai/tablevqabench/blob/master/src/example_prompts.py (VWTQ_PROMPT L8-L19)",
"notes": "Tier 1: the authors' official per-split few-shot vision prompt, deployed verbatim (jinja-ified {question} -> {{ question }}; <image> placeholder prepended). Restructured from one 'main' subset into one subset per official split on 2026-07-07 so each split renders its official prompt."
}
},
"vwtq_syn": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "short_answer_qa",
"score_pipeline": [
"exact-match",
"rule-match"
],
"score_protocol": {
"reference": "Official naver-ai/tablevqabench src/evaluate_tablevqa_dataset.py:17-33 (per-split dispatch) — vwtq/vwtq_syn use WikiTableQuestions denotation matching (to_value_list/check_denotation with number/date normalization); vtabfact scores lowercase 'true'/'false' containment in the prediction (evaluate_tabfact, lines 45-64); fintabnetqa uses normalized exact match with $/unit-word stripping ('relieved accuracy'). Ported in VLMEvalKit@vlmeval/dataset/image_vqa.py:2366-2396 + vlmeval/dataset/utils/tablevqabench.py:54-158. Headline = mean accuracy per split.",
"note": "Fully deterministic (rule) but with per-split normalization the generic exact/template chain will NOT reproduce: WTQ denotation normalization and fintabnet unit stripping. vtabfact gt is stored as 'true'/'false' (mapped from source gt '1'/'0'), matching what the official evaluator expects from a generative model: evaluate_tabfact (src/evaluate_tablevqa_dataset.py:56-64) lowercases the prediction and scores \"'true' in pred and gt==1\" / \"'false' in pred and gt==0\" as correct, and the official VTABFACT_PROMPT (src/example_prompts.py:21-29) instructs 'You should only answer true or false.'. Official also uses split-specific prompts (VTABFACT_PROMPT etc., image_vqa.py:2398-2414) whereas the published prompt_template is the generic short-answer prompt. Official reporting is per-split accuracy averaged over 4 splits; published copy stores splits as HF splits under one subset."
},
"prompt_template": "<image>You are asked to answer questions asked on an image.\nYou should answer the question with a single word.\nExample: \nQuestion: what was the only year mr. wu competed in the olympic games?\nAnswer: 2004\nQuestion: which township in pope county, arkansas has the least amount of water area?\nAnswer: Freeman\nIf you have multiple answers, please separate them with || marks. Example: Apple||Banana||Tomato\n\nQuestion: {{ question }}\nAnswer:",
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "qa_id"
},
"question": {
"from": "question"
},
"answer": {
"from": "gt",
"optional": true,
"note": "vtabfact split only: source gt '1'/'0' mapped to 'true'/'false' per the official evaluator (naver-ai/tablevqabench src/evaluate_tablevqa_dataset.py:56-64: lowercase 'true'/'false' containment vs gt 1/0; src/example_prompts.py:21-29: 'You should only answer true or false.'). vwtq/vwtq_syn/fintabnetqa gts are verbatim source values."
},
"source": {
"format": "huggingface",
"url": {
"vwtq_syn": "https://huggingface.co/datasets/terryoo/TableVQA-Bench"
}
}
},
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/naver-ai/tablevqabench/blob/master/src/example_prompts.py (VWTQ_PROMPT L8-L19 (vwtq_syn is the synthetic-style WTQ variant, same official prompt))",
"notes": "Tier 1: the authors' official per-split few-shot vision prompt, deployed verbatim (jinja-ified {question} -> {{ question }}; <image> placeholder prepended). Restructured from one 'main' subset into one subset per official split on 2026-07-07 so each split renders its official prompt."
}
},
"vtabfact": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "yes_no_qa",
"score_pipeline": [
"exact-match",
"rule-match"
],
"score_protocol": {
"reference": "Official naver-ai/tablevqabench src/evaluate_tablevqa_dataset.py:17-33 (per-split dispatch) — vwtq/vwtq_syn use WikiTableQuestions denotation matching (to_value_list/check_denotation with number/date normalization); vtabfact scores lowercase 'true'/'false' containment in the prediction (evaluate_tabfact, lines 45-64); fintabnetqa uses normalized exact match with $/unit-word stripping ('relieved accuracy'). Ported in VLMEvalKit@vlmeval/dataset/image_vqa.py:2366-2396 + vlmeval/dataset/utils/tablevqabench.py:54-158. Headline = mean accuracy per split.",
"note": "Fully deterministic (rule) but with per-split normalization the generic exact/template chain will NOT reproduce: WTQ denotation normalization and fintabnet unit stripping. vtabfact gt is stored as 'true'/'false' (mapped from source gt '1'/'0'), matching what the official evaluator expects from a generative model: evaluate_tabfact (src/evaluate_tablevqa_dataset.py:56-64) lowercases the prediction and scores \"'true' in pred and gt==1\" / \"'false' in pred and gt==0\" as correct, and the official VTABFACT_PROMPT (src/example_prompts.py:21-29) instructs 'You should only answer true or false.'. Official also uses split-specific prompts (VTABFACT_PROMPT etc., image_vqa.py:2398-2414) whereas the published prompt_template is the generic short-answer prompt. Official reporting is per-split accuracy averaged over 4 splits; published copy stores splits as HF splits under one subset."
},
"prompt_template": "<image>You are asked to answer whether the statement is True or False based on given image\nYou should only answer true or false.\nExample: \nStatement: the milwaukee buck win 6 game in the 2010 - 11 season\nAnswer: True\nStatement: only the top team score above the average of 8.8\nAnswer: False\n\nStatement: {{ question }}\nAnswer:",
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "qa_id"
},
"question": {
"from": "question"
},
"answer": {
"from": "gt",
"optional": true,
"note": "vtabfact split only: source gt '1'/'0' mapped to 'true'/'false' per the official evaluator (naver-ai/tablevqabench src/evaluate_tablevqa_dataset.py:56-64: lowercase 'true'/'false' containment vs gt 1/0; src/example_prompts.py:21-29: 'You should only answer true or false.'). vwtq/vwtq_syn/fintabnetqa gts are verbatim source values."
},
"source": {
"format": "huggingface",
"url": {
"vtabfact": "https://huggingface.co/datasets/terryoo/TableVQA-Bench"
}
}
},
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/naver-ai/tablevqabench/blob/master/src/example_prompts.py (VTABFACT_PROMPT L21-L31)",
"notes": "Tier 1: the authors' official per-split few-shot vision prompt, deployed verbatim (jinja-ified {question} -> {{ question }}; <image> placeholder prepended). Restructured from one 'main' subset into one subset per official split on 2026-07-07 so each split renders its official prompt."
}
},
"fintabnetqa": {
"language": [
"en"
],
"modalities": [
"single_image_start"
],
"task_type": "short_answer_qa",
"score_pipeline": [
"exact-match",
"rule-match"
],
"score_protocol": {
"reference": "Official naver-ai/tablevqabench src/evaluate_tablevqa_dataset.py:17-33 (per-split dispatch) — vwtq/vwtq_syn use WikiTableQuestions denotation matching (to_value_list/check_denotation with number/date normalization); vtabfact scores lowercase 'true'/'false' containment in the prediction (evaluate_tabfact, lines 45-64); fintabnetqa uses normalized exact match with $/unit-word stripping ('relieved accuracy'). Ported in VLMEvalKit@vlmeval/dataset/image_vqa.py:2366-2396 + vlmeval/dataset/utils/tablevqabench.py:54-158. Headline = mean accuracy per split.",
"note": "Fully deterministic (rule) but with per-split normalization the generic exact/template chain will NOT reproduce: WTQ denotation normalization and fintabnet unit stripping. vtabfact gt is stored as 'true'/'false' (mapped from source gt '1'/'0'), matching what the official evaluator expects from a generative model: evaluate_tabfact (src/evaluate_tablevqa_dataset.py:56-64) lowercases the prediction and scores \"'true' in pred and gt==1\" / \"'false' in pred and gt==0\" as correct, and the official VTABFACT_PROMPT (src/example_prompts.py:21-29) instructs 'You should only answer true or false.'. Official also uses split-specific prompts (VTABFACT_PROMPT etc., image_vqa.py:2398-2414) whereas the published prompt_template is the generic short-answer prompt. Official reporting is per-split accuracy averaged over 4 splits; published copy stores splits as HF splits under one subset."
},
"prompt_template": "<image>You are asked to answer questions asked on a image.\nYou should answer the question within a single word or few words.\nIf units can be known, the answer should include units such as $, %, million and etc.\nExample: \nQuestion: What were the total financing originations for the fiscal year ended October 31, 2004?\nAnswer: $3,852 million\nQuestion: What is the time period represented in the table?\nAnswer: October 31\nQuestion: What was the percentage of net sales for selling, general and administrative expenses in 2006?\nAnswer: 34.2%\nQuestion: {{ question }}\nAnswer:",
"mapping_from_source": {
"media": {
"from": "image",
"type": "list",
"min_items": 1,
"max_items": 1
},
"id": {
"from": "qa_id"
},
"question": {
"from": "question"
},
"answer": {
"from": "gt",
"optional": true,
"note": "vtabfact split only: source gt '1'/'0' mapped to 'true'/'false' per the official evaluator (naver-ai/tablevqabench src/evaluate_tablevqa_dataset.py:56-64: lowercase 'true'/'false' containment vs gt 1/0; src/example_prompts.py:21-29: 'You should only answer true or false.'). vwtq/vwtq_syn/fintabnetqa gts are verbatim source values."
},
"source": {
"format": "huggingface",
"url": {
"fintabnetqa": "https://huggingface.co/datasets/terryoo/TableVQA-Bench"
}
}
},
"prompt_template_source": {
"origin": "official",
"reference": "https://github.com/naver-ai/tablevqabench/blob/master/src/example_prompts.py (FINTABNETQA_PROMPT L33-L46)",
"notes": "Tier 1: the authors' official per-split few-shot vision prompt, deployed verbatim (jinja-ified {question} -> {{ question }}; <image> placeholder prepended). Restructured from one 'main' subset into one subset per official split on 2026-07-07 so each split renders its official prompt."
}
}
}
} |