rasmbenchanon commited on
Commit
b018641
·
verified ·
1 Parent(s): d9f0264

Shorten dataset card

Browse files
Files changed (1) hide show
  1. README.md +31 -61
README.md CHANGED
@@ -68,73 +68,45 @@ configs:
68
 
69
  # RasmBench
70
 
71
- RasmBench is a benchmark of 1,002 rendered Arabic documents that contain charts and tables. Each document has six questions, one per level of Bloom's taxonomy (Remember, Understand, Apply, Analyze, Evaluate, Create), asked both as multiple choice and as open-ended questions: 6,012 items per format.
72
 
73
- The documents are built from tables in Arabic Wikipedia and come in three value regimes, 334 documents each:
 
 
74
 
75
- - **factual**: the figure shows the real values from the source table.
76
- - **counterfactual**: every number is perturbed so that the figure contradicts the real-world value, and reader-checkable constraints (totals, percentages, the order of years) are then restored. Labels stay real.
77
- - **neutral**: labels are replaced with placeholders and numbers with random values of similar magnitude, so there is no link to world knowledge.
78
-
79
- Pages are rendered at five degradation tiers: A born-digital, B print scan, C complex-layout scan, D camera capture, E historical print.
80
-
81
- > **Counterfactual items contain deliberately false statistics about real entities.** They are marked `values_altered = true`. Do not use them as factual data or train on them as real data.
82
 
83
  ## Configurations
84
 
85
- | Config | Splits (rows) | What it is |
86
  |---|---|---|
87
- | `documents` (default) | `test` (1,002) | The benchmark as released: page image, clean image, six questions |
88
- | `paired_rerenders` | `factual_values` (668), `counterfactual_values` (668) | Every factual and counterfactual document drawn with both value sets, with questions re-keyed to the values shown |
89
- | `figure_as_text` | `image_table` (100), `image_table_control` (50), `table_only` (100), `own_table` (230) | The figure's data given as a markdown table: with the image, without it, and as the model's own transcription (5 models × 46 documents) |
90
- | `translation` | `ar_doc_ar_q`, `ar_doc_en_q`, `en_doc_ar_q`, `en_doc_en_q` (98 each) | Arabic or English document × Arabic or English question |
91
- | `chart_to_table` | `selection` (46), `pairs` (97), `tiers` (245) | Transcribe the single figure as a table; the reference is what the chart actually draws |
92
- | `model_outputs` | `test` (201,593) | Every model answer in every condition, with judge scores for open-ended answers |
93
- | `second_judge` | `test` (4,252) | A second judge (mistral-medium-3) on 299 + 299 sampled open-ended questions |
94
-
95
- The diagnostic conditions use the 100 selected documents (the 50 with the highest and 50 with the lowest mean score across the ten open models) and a control set of 50 random other documents. Both lists are in `metadata/` and in the `selection` column of `documents` (`easy`, `hard`, `control`).
96
 
97
- ## Loading
98
 
99
  ```python
100
  from datasets import load_dataset
101
-
102
  docs = load_dataset("rasmbench/rasmbench_revised", split="test")
103
- pairs = load_dataset("rasmbench/rasmbench_revised", "paired_rerenders")
104
  outputs = load_dataset("rasmbench/rasmbench_revised", "model_outputs", split="test")
105
  ```
106
 
107
- ## Fields (`documents`)
108
-
109
- | Field | Description |
110
- |---|---|
111
- | `uid` | `<regime>/<doc_id>`. **Key on `uid`, not `doc_id`**: 21 `doc_id`s occur in more than one regime |
112
- | `regime` | `factual`, `counterfactual` or `neutral` |
113
- | `values_altered` | `true` for counterfactual documents |
114
- | `tier`, `tier_name` | Degradation tier applied to the page (A–E) |
115
- | `template_nominal_tier` | The tier the template was designed for; differs from `tier` on 168 documents |
116
- | `degradation` | Degradation operations applied, with parameters |
117
- | `composition` | `1 chart only`, `1 table only` or `mixed (chart + table)` |
118
- | `anchor`, `doc_type`, `title` | Document-level anchor, document type, source title |
119
- | `n_charts`, `n_tables` | Number of figures (at most two) |
120
- | `selection` | `easy` / `hard` (100 selected), `control`, or null |
121
- | `image` | The page as evaluated |
122
- | `image_clean` | The same page before degradation (not used in the paper's evaluation) |
123
- | `questions` | Six questions: `level`, `question`, `options` (4), `answer_index` (0–3), `anchor` (`chart`, `table` or `text+figure`), `distractor_pool` |
124
 
125
- `distractor_pool` holds the candidates the distractors were chosen from. **Remove it before prompting a model**: it leaks the answer space.
 
 
 
 
126
 
127
- The open-ended form of each question is the same `question` without options; its reference answer is `options[answer_index]`.
128
 
129
- Variant configs keep the record format of `documents`. `uid` starts with the condition name, and `source_uid` points to the released document. `values_shown` gives the value set the image draws. In `paired_rerenders`, questions carry `rekeyed` and `evidence` (a JSON string) for the re-keying step, and `in_paired_150` / `in_selected_68` mark the subsets used in the paper.
130
-
131
- ## Fields (`model_outputs`)
132
-
133
- One row per model × condition × item × format. `condition` is `original` for the released pages, otherwise the variant name. `document_uid` joins to the `uid` of the variant config. `protocol` is `mcq` or `openended`. For MCQ, `predicted_index` is the parsed option (null if no letter could be parsed; the paper counts it as wrong) and `correct` is 0/1. Open-ended answers have `judge_model`, `judge_score` (0–5), `judge_label` and, for later runs, `judge_reason`. The original open-ended answers were judged by gemini-3.5-flash and all later answers by gemini-3.8-flash, so compare open-ended scores only within one `judge_model`. Chart-to-table answers are scored with RMS-F1, not a judge.
134
-
135
- ## Evaluation prompts
136
-
137
- MCQ (the four options follow as `A) …` to `D) …`):
138
 
139
  ```
140
  اقرأ الوثيقة في الصورة وأجب عن السؤال التالي.
@@ -154,21 +126,19 @@ Open-ended:
154
  {question}
155
  ```
156
 
157
- Decoding was greedy, with at most 128 new tokens for MCQ, 256 for open-ended answers and 1,024 for chart-to-table transcription. The judge rubric and the chart-to-table prompts are in the paper's appendix.
158
-
159
- ## Known limitations
160
 
161
- - Questions were written by gemini-2.5-flash-lite, and open-ended answers are judged by Gemini models, so items and scores may favour models that reason like Gemini.
162
- - Questions of re-rendered documents were re-keyed automatically. In a hand check of 50, between 1 and 3 were wrong.
163
- - English documents are machine translations (gemini-2.5-flash, failures retried with gemini-2.5-pro). Two of the 100 selected documents failed both and are left out (`metadata/translation_failed.json`). Qwen3.6-27B and gemma-4-31B answered 76 of the 98 translated documents.
164
- - Tiers were not assigned at random, so comparisons between tiers are observational.
165
- - Documents are synthetic renders. The rendering engine is not part of this release.
166
 
167
- ## Licence and attribution
 
 
 
 
168
 
169
- All content is released under [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/). The source tables come from Arabic Wikipedia (CC BY-SA 4.0); each document's `title` names its source table's subject. If you redistribute or adapt the data, keep the same licence and credit Arabic Wikipedia and this dataset.
170
 
171
- Model-generated content: the questions, translations and judge scores were produced with Gemini models, and the answers in `model_outputs` with the listed open-weight models and mistral-medium-3. Uses of those outputs may also be subject to the providers' terms.
172
 
173
  ## Citation
174
 
 
68
 
69
  # RasmBench
70
 
71
+ 1,002 rendered Arabic documents with charts and tables, each with six questions (one per Bloom level) in multiple-choice and open-ended form: 6,012 items per format. Documents are built from Arabic Wikipedia tables in three value regimes (334 each) and five degradation tiers (A born-digital, B print scan, C complex-layout scan, D camera capture, E historical print).
72
 
73
+ - **factual**: the real values.
74
+ - **counterfactual**: every number perturbed, with totals, percentages and year order restored. Labels stay real.
75
+ - **neutral**: placeholder labels and random values.
76
 
77
+ > **Counterfactual items contain deliberately false statistics about real entities** (`values_altered = true`). Do not use them as factual data or train on them as real data.
 
 
 
 
 
 
78
 
79
  ## Configurations
80
 
81
+ | Config | Splits (rows) | Contents |
82
  |---|---|---|
83
+ | `documents` (default) | `test` (1,002) | The benchmark: page image, undegraded image, six questions |
84
+ | `paired_rerenders` | `factual_values`, `counterfactual_values` (668 each) | Each factual/counterfactual document drawn with both value sets, questions re-keyed |
85
+ | `figure_as_text` | `image_table` (100), `image_table_control` (50), `table_only` (100), `own_table` (230) | Figure data as a markdown table: with the image, without it, or as the model's own transcription |
86
+ | `translation` | `ar_doc_ar_q`, `ar_doc_en_q`, `en_doc_ar_q`, `en_doc_en_q` (98 each) | Arabic/English document × Arabic/English question |
87
+ | `chart_to_table` | `selection` (46), `pairs` (97), `tiers` (245) | Transcribe the figure; the reference is what the chart draws |
88
+ | `model_outputs` | `test` (201,593) | Every model answer in every condition, with judge scores |
89
+ | `second_judge` | `test` (4,252) | mistral-medium-3 as a second judge on 598 sampled questions |
 
 
90
 
91
+ Diagnostic conditions use 100 selected documents (the 50 easiest and 50 hardest across ten models) and 50 random control documents, listed in `metadata/` and in the `selection` column. Full page annotations are in `annotations/` (re-rendered variants also carry theirs in an `annotation` column).
92
 
93
  ```python
94
  from datasets import load_dataset
 
95
  docs = load_dataset("rasmbench/rasmbench_revised", split="test")
 
96
  outputs = load_dataset("rasmbench/rasmbench_revised", "model_outputs", split="test")
97
  ```
98
 
99
+ ## Notes
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
100
 
101
+ - **Key on `uid`** (`<regime>/<doc_id>`), not `doc_id`: 21 `doc_id`s occur in more than one regime. In variant configs `uid` starts with the condition name, and `source_uid` points to the released document.
102
+ - **Drop `distractor_pool` before prompting**; it leaks the answer space.
103
+ - The open-ended reference answer is `options[answer_index]`.
104
+ - `tier` is the degradation applied. `template_nominal_tier` is the template's design tier and differs on 168 documents.
105
+ - `model_outputs`: `condition` is `original` or the variant name, and `document_uid` joins to the variant's `uid`. Unparsed MCQ answers have a null `predicted_index` and count as wrong. Original open-ended answers were judged by gemini-3.5-flash and later ones by gemini-3.8-flash, so compare scores within one `judge_model`.
106
 
107
+ ## Prompts
108
 
109
+ MCQ, with options as `A) …` to `D) …`:
 
 
 
 
 
 
 
 
110
 
111
  ```
112
  اقرأ الوثيقة في الصورة وأجب عن السؤال التالي.
 
126
  {question}
127
  ```
128
 
129
+ Greedy decoding with at most 128 new tokens for MCQ, 256 for open-ended answers and 1,024 for chart-to-table transcription. The judge rubric and chart-to-table prompts are in the paper's appendix.
 
 
130
 
131
+ ## Limitations
 
 
 
 
132
 
133
+ - Questions were written by gemini-2.5-flash-lite and open-ended answers judged by Gemini, which may favour Gemini-like models.
134
+ - Re-rendered questions were re-keyed automatically; between 1 and 3 of 50 hand-checked were wrong.
135
+ - English documents are machine translations. 2 of the 100 selected failed to translate (`metadata/translation_failed.json`), and Qwen3.6-27B, gemma-4-31B, Qwen3-VL-4B and Qwen3.5-9B answered 76 of the 98.
136
+ - Tiers were not assigned at random, so tier comparisons are observational.
137
+ - The rendering engine is not part of this release.
138
 
139
+ ## Licence
140
 
141
+ CC BY-SA 4.0, derived from Arabic Wikipedia (CC BY-SA 4.0). Credit Arabic Wikipedia and this dataset, and keep the licence. Questions, translations and judge scores were generated with Gemini models and answers with the listed models, so their providers' terms may also apply.
142
 
143
  ## Citation
144