mhfisher/CaTSBench

CaTS-Bench Dataset

A comprehensive benchmark for evaluating multi-modal models on time series understanding, captioning, and reasoning tasks across diverse domains.

Quickstart

Install

# To unpack the dataset archives in-place (each <name>.zip -> <name>/)
python unpack.py
# To use codebase
git clone https://github.com/LuckerZOfficiaL/CaTS-Bench.git 
pip install -r requirements.txt

Run and evaluate

Run inference on a pre-trained model:

# Text + Image evaluation (multimodal)
python -m source.inference.llama_infer

# Other supported models
python -m source.inference.internvl_infer
python -m source.inference.qwenvl_infer
python -m source.inference.phi_infer

Score model performance:

# Evaluate on the small benchmark (460 tasks)
python source/score_unified.py QA_hard_small/tasks.json your_model_output/ --output results.json

# Evaluate on full benchmark (~38k tasks)
python source/score_unified.py QA/tasks.json your_model_output/ --output full_results.json

Generate new captions for time series data:

python source/generate_captions_baseline.py

Task Types & Examples

The dataset includes 14 distinct task types testing different aspects of time series understanding:

Retrieval tasks (text + image or text-only):

1. caption_retrieval_same_domain - Match time series to its correct caption (same-domain distractors)
2. caption_retrieval_cross_domain - Match time series to its correct caption (cross-domain distractors)
3. caption_retrieval_perturbed - Caption retrieval with numerically perturbed distractors
4. ts_retrieval_same_domain - Find the time series matching a caption (same-domain distractors)
5. ts_retrieval_cross_domain - Find the time series matching a caption (cross-domain distractors)
6. ts_retrieval_perturbed - TS retrieval with numerically perturbed distractors
7. plot_retrieval_same_domain - Visual matching of plot to caption (same-domain distractors)
8. plot_retrieval_cross_domain - Visual matching of plot to caption (cross-domain distractors)
9. paraphrase_consistency - Verify consistent understanding across paraphrased captions

Comparison tasks (text-only): 10. ts_comparison_amplitude - Compare amplitude between two time series 11. ts_comparison_mean - Compare mean values between two time series 12. ts_comparison_volatility - Compare volatility between two time series 13. ts_comparison_peak_earlier - Identify which series peaks earlier 14. ts_comparison_same_phenomenon - Determine if two series describe the same phenomenon

All retrieval tasks support both multimodal (text + image) and text-only evaluation modes.

Dataset Structure

Train Data (15,995 samples)

train_data/ contains:

• synth_gt_captions/ - Synthetically generated ground-truth captions (one .txt per time series)
• metadata/ - Contextual and statistical metadata for each time series (.json per series)
• plots/ - Visual plot images (JPEG format)
• time series/ - Numerical time series values (.txt format, one value per line)

Note: All files share aligned IDs across subfolders — e.g., agriculture_0_train.txt, agriculture_0_train.json, and agriculture_0_train.jpeg all correspond to the same time series.

Test Data (4,005 samples)

test_data/ contains:

Note: All files share aligned IDs across subfolders — e.g., agriculture_0_test.txt, agriculture_0_test.json, and agriculture_0_test.jpeg all correspond to the same time series. This alignment holds across synth_gt_captions, human_rewritten_captions, metadata, plots, and time series.

• synth_gt_captions/ - Synthetically generated ground-truth captions for all 4,005 test series
• human_rewritten_captions/ - Gold-standard human-rewritten captions for 1,746 test series — these are the most reliable and valuable ground-truth captions in the dataset, manually verified for factual accuracy
• synth_gt_captions_gemma_paraphrased/ - Gemma paraphrases of GT captions (549 samples)
• synth_gt_captions_GPT-4o_paraphrased/ - GPT-4o paraphrases of GT captions (549 samples)
• synth_gt_captions_llama_paraphrased/ - Llama paraphrases of GT captions (549 samples)
• metadata/ - Contextual and statistical metadata for each time series
• plots/ - Visual plot images (JPEG format)
• time series/ - Numerical time series values
• config.yaml - Configuration used to generate this test split

Note: The 1,746 series with human-rewritten captions form the primary gold evaluation set and should be preferred over synthetic captions for rigorous evaluation.

Task Files

• QA/tasks.json - Complete benchmark (38,402 tasks)
• QA_hard_small/tasks.json - Curated difficult subset (460 tasks, used in paper)
• QA_hard_medium/tasks.json - Medium difficult subset (~4k tasks)
• QA_hard/tasks.json - Extended difficult subset (~7k tasks)
• QA/temporal_matching/ - Additional temporal matching tasks (JSONL format)

Caption Prompts

caption_prompts/ contains original prompts used for caption generation:

• train/prompts.json - Training set prompts
• test/prompts.json - Test set prompts

Task Format

Most tasks in tasks.json follow this structure:

{
  "task_id": "caption_retrieval_same_domain_agriculture_0_test",
  "task_type": "caption_retrieval_same_domain",
  "ts_name": "agriculture_0_test",
  "prompt_with_image": "Multi-modal prompt text (time series values + image reference)",
  "prompt_no_image": "Text-only prompt text (time series values only)",
  "ground_truth": "B",
  "image_paths": ["agriculture_0_test.jpeg"]
}

ts_comparison tasks use a single prompt field instead of prompt_with_image/prompt_no_image.

Domain Coverage

Time series data spans 11 real-world domains:

• Agriculture
• Air Quality
• Border Crossing
• CO2 Emissions
• COVID-19
• Crime
• Demography
• Diet
• Online Retail
• Road Injuries
• Walmart (retail sales)

Evaluation Metrics

• Accuracy - Primary metric across all task types
• Domain-specific Performance - Breakdown by data domain
• Task-type Analysis - Performance across different reasoning types
• Multimodal vs Text-only - Comparison of modality effects