pretty_name: Comprehensive Multimodal Decision Benchmark 1500
license: other
license_name: source-dependent
license_link: https://huggingface.co/datasets/SurdAI/CMDB-1500/blob/main/LICENSES.md
language:
- en
- zh
size_categories:
- 1K<n<10K
task_categories:
- text-classification
- question-answering
- visual-question-answering
tags:
- benchmark
- decision-models
- multimodal
- calibration
- multi-select
- text
- image
configs:
- config_name: default
data_files:
- split: test
path: data/test.parquet
CMDB-1500
Comprehensive Multimodal Decision Benchmark 1500, curated by SurdAI, contains 1,500 candidate-based decision questions: 1,200 text-only and 300 image-based, spanning 10 domains and 79 sources and subtasks.
The benchmark covers answer selection, action selection, binary judgments, ordinal ratings, and multi-select decisions in a common format.
Interactive leaderboard / 交互榜单 · Benchmark introduction / 基准介绍 · Example questions · Source catalog
Results and models
Explore the interactive CMDB-1500 leaderboard on the Surd AI website. It separates text (1,200 questions), image (300 questions), and full-benchmark (1,500 questions) accuracy, with probability calibration reported separately. Text-only models do not receive an image or full-benchmark score.
Selected Surd AI results use effort 2 and the same 1,500-question benchmark:
| Model | Text accuracy ↑ | Image accuracy ↑ | Full-benchmark accuracy ↑ |
|---|---|---|---|
| SPX-CD-Pro | 79.33% | 88.67% | 81.20% |
| SPX-CD-Flash | 77.58% | 87.33% | 79.53% |
| SPX-CD-Omni | 74.67% | 82.67% | 76.27% |
These are team-run evaluations, not official upstream leaderboard submissions. SPX-CD-Omni's reported results use the released LoRA adapter with an AWQ INT4 base; they are not measurements of the default BF16 example configuration. The model card and evaluation settings describe the configuration and metric scopes.
The website's cross-model I/C comparison uses a shared text-probability cohort; it is distinct from the model card's benchmark-specific I/C scopes. Compare I/C values only within the same scoring scope. See the SPX-CD release and API overview for Pro and Flash.
Coverage
| Domain | Questions | Tasks |
|---|---|---|
| Language, knowledge, and judgment | 450 | Science, intent classification, inference, evidence, paraphrase similarity, and response assessment |
| Professional and applied decisions | 300 | Engineering, agents, commerce, finance, law, data, documents, product, support, and safety |
| Multimodal decisions | 300 | Diagram understanding, visual reasoning, science questions, image classification, and visual question answering |
| Games | 100 | Static game states and action selection across seven subtasks |
| Multi-select | 100 | MultiRC and GoEmotions |
| Safety and adversarial decisions | 100 | Content safety, phishing, and adversarial instruction handling |
| Embodied action selection | 50 | ALFRED, ALFWorld, and VirtualHome |
| Chinese ASR hypothesis selection | 50 | Selection among text transcription hypotheses |
| Tool-call necessity | 25 | BFCL-derived judgments of whether a tool call is needed |
| Spatial and complex rules | 25 | Spatial relations and rule-based decisions |
| Total | 1,500 | 1,200 text-only + 300 image-based |
The image subset contains 322 images; some questions use more than one image.
| Record format | kind |
Questions | Expected prediction |
|---|---|---|---|
| Choice | choice |
1,077 | One candidate key |
| Ordered rating | score |
166 | One key on an ordered scale |
| Binary format | noul |
157 | One key from the two supplied candidates |
| Multi-select | multi_choice |
100 | A set of candidate keys |
Counts follow the kind field. choice includes two-option questions; multi_choice permits one or more correct answers. Ordered ratings cover helpfulness, verbosity, similarity, hate-speech levels, and game-state assessments. Rating direction follows the question.
Sources
- Language and judgment: 19 JevBench configurations, including Banking77, ARC-Challenge, BoolQ, CLINC, MASSIVE, MMLU, ChaosNLI, MNLI, SST-5, and HelpSteer2.
- Professional decisions: 31 Atlan Decision Bench tasks covering engineering, agents, commerce, finance, law, data, documents, product, support, and safety.
- Multimodal decisions: 30 questions each from AI2D, Fashion200k, GQA, MMMU, MMMU-Pro, MathVista, ScienceQA, TextVQA, VQAv2, and VizWiz.
- Other domains: Open-Jev games, MultiRC, GoEmotions, ChineseHP, Aegis, PhishNChips, JevAdvBench, BFCL v3, THIS-THAT, ALFRED, ALFWorld, and VirtualHome.
SOURCES.md lists all question counts, upstream links, and task definitions. Subtasks may share an upstream dataset.
Load the dataset
from datasets import load_dataset
bench = load_dataset("SurdAI/CMDB-1500", split="test")
item = bench[0]
model_input = {
key: item[key]
for key in ("state", "question", "options", "kind", "images")
}
Data format
The test split is available as Parquet with embedded images and JSONL. JSONL image paths are relative to the dataset directory.
| Fields | Contents |
|---|---|
state, state_format |
Context as text or serialized JSON |
question, options |
Decision instruction and candidate keys/texts |
kind, modality, images |
Output type and image inputs |
correct_keys, target_probs |
Reference answers and optional auxiliary targets |
id, source, category |
Question identifier, source, and domain |
source_id, original_split, source_metadata, label_origin |
Source provenance and reference-label basis |
Parse state when state_format is json. target_probs and source_metadata are JSON-serialized strings. Candidate keys are specific to each question. Use the model-input fields shown above; reference answers and provenance are for evaluation.
Evaluation
Use accuracy for categorical choices, including binary questions, exact-set match and set F1 for multi-select, and accuracy plus level error for ordinal ratings. Calibration metrics use valid model probability outputs and the stated reference labels; report the included sample count and the binning rule. Auxiliary target distributions are not model confidence predictions.
Report domain scores and text/image results alongside an overall score, specifying the aggregation weights. Retain the full scope denominator when a native model interface cannot represent a question; report such failures separately. Text-only evaluations cover 1,200 questions and must not be presented as full 1,500-question results.
These are fixed-input decisions. Game and embodied questions assess state/action selection; BFCL questions assess tool-call necessity; Chinese ASR questions use text hypotheses. GQA, VQAv2, TextVQA, and VizWiz use yes/no subsets. Scores describe these selected tasks rather than complete upstream benchmark evaluations.
Attribution and licensing
LICENSES.md contains source-specific terms and required attribution, including HelpSteer2. ChaosNLI is CC BY-NC 4.0, with attribution and noncommercial conditions; ScienceQA also has noncommercial conditions. MultiNLI permits modification and redistribution for typical machine-learning uses, with underlying source terms retained. An explicit redistribution license for SST-5 has not been identified in the inspected original release documentation. MathVista permits test-set use only. Upstream terms apply to the corresponding records.
Please cite CMDB-1500, SurdAI, together with the upstream datasets used.