File size: 6,136 Bytes
c9284d2
 
 
 
 
f4e8e56
 
 
 
 
 
 
 
 
 
160b9a0
f4e8e56
 
 
160b9a0
f4e8e56
 
 
160b9a0
f4e8e56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cbf2b83
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0fbcad6
 
 
cbf2b83
0fbcad6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cbf2b83
 
0fbcad6
cbf2b83
0fbcad6
cbf2b83
 
 
0fbcad6
 
cbf2b83
 
 
 
0fbcad6
 
 
 
cbf2b83
f4e8e56
 
 
 
 
160b9a0
 
 
f4e8e56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dc4fedf
 
 
 
 
f4e8e56
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# HiTSR Dataset

[![arXiv](https://img.shields.io/badge/arXiv-2604.17295-FF6B6B.svg)]([https://arxiv.org/abs/2504.XXXXX](https://arxiv.org/abs/2604.17295))

This is the official dataset repository of the ACL 2026 Findings paper: "LLaTiSA: Towards Difficulty-Stratified Time Series Reasoning from Visual Perception to Semantics".

A comprehensive **multimodal time series understanding and reasoning dataset** with multiple complexity levels.

## Overview

This dataset contains time series data paired with visual representations and natural language instructions for time series analysis tasks. The dataset is organized into **3 levels of complexity** with corresponding train/test splits.

### Dataset Statistics

- **Level 1 (Basic)**: Single series analysis (min/max detection, trend analysis)
  - Training samples: 54,000
  - Test samples: Multiple variants (minmax, multiseries, startend, subseries)

- **Level 2 (Intermediate)**: Multi-series analysis and relationships
  - Training samples: 45,632
  - Test categories: global, local, numerical

- **Level 3 (Advanced)**: Complex reasoning and annotations
  - Training samples: 3,515
  - Test samples: Final test set

## Data Structure

Each sample contains:

```json
{
  "id": 1,
  "timeseries": [[float_values]],
  "prompt": "Multi-modal prompt with image tags",
  "answer": "Model answer to the question",
  "2img_prompt": "Detailed instructions for image interpretation",
  "prompt_1": "Variant 1 of the question",
  "prompt_2": "Variant 2 of the question",
  "answer_1": "Answer in format 1",
  "answer_2": "Answer in format 2",
  "images": ["image_url_1", "image_url_2"]
}
```

### Key Fields

- **id**: Unique identifier for each sample
- **timeseries**: Array of time series values (floats)
- **prompt**: Main question/instruction with image references (e.g., `<image>`)
- **answer**: Expected model response
- **2img_prompt**: Detailed instructions for interpreting high-density numeric grids
- **prompt_1/prompt_2**: Alternative question formats
- **answer_1/answer_2**: Alternative answer formats
- **images**: URLs to corresponding plot and numeric grid images

## Image Generation

The `images` field in each sample contains URLs to visual representations of the time series data. To generate these images and populate the `images` field:

1. **Clone the data conversion repository:**
   ```bash
   git clone https://github.com/RainingNovember/LLaTiSA.git
   cd LLaTiSA/data_convert
   ```

2. **Install required dependencies:**
   ```bash
   pip install -r requirements.txt
   ```

3. **Run the appropriate conversion script based on dataset level:**
   
   **For Level 1 datasets:**
   ```bash
   python data_convert_l1.py \
     --input /path/to/level1_train.json \
     --output /path/to/level1_train_with_images.json \
     --plot_dir ./images/plots \
     --num_dir ./images/numeric \
     --plot_prefix "https://your-hosting.com/images/plots" \
     --num_prefix "https://your-hosting.com/images/numeric" \
     --sample_ratio 1.0
   ```
   
   **For Level 2 datasets:**
   ```bash
   python data_convert_l2.py \
     --input /path/to/level2_train.json \
     --output /path/to/level2_train_with_images.json \
     --plot_dir ./images/plots \
     --num_dir ./images/numeric \
     --plot_prefix "https://your-hosting.com/images/plots" \
     --num_prefix "https://your-hosting.com/images/numeric" \
     --sample_ratio 1.0
   ```
   
   **For Level 3 datasets:**
   ```bash
   python data_convert_l3.py \
     --input /path/to/level3_train.json \
     --output /path/to/level3_train_with_images.json \
     --plot_dir ./images/plots \
     --num_dir ./images/numeric \
     --plot_prefix "https://your-hosting.com/images/plots" \
     --num_prefix "https://your-hosting.com/images/numeric" \
     --sample_ratio 1.0
   ```

4. **Upload generated images to a hosting service** (e.g., GitHub, Imgur, or cloud storage) and update the URL prefixes in the commands above.

5. **The output JSON files will have the `images` field populated** with the correct URLs:
   ```json
   {
     "images": [
       "https://your-hosting.com/images/plots/plot_1.png",
       "https://your-hosting.com/images/numeric/num_1.png"
     ]
   }
   ```

**Notes:**
- Each script generates two types of images per sample: trend plots (`plot_*.png`) and high-density numeric grids (`num_*.png`)
- Use `--sample_ratio 1.0` to process all samples (default is 0.5)
- The scripts automatically update the `prompt` and `answer` fields based on the dataset level requirements

## File Organization

```
AAA-HiTSR/
├── Train/
│   ├── l1_train_llatisa.json      (Level 1: 54,000 samples)
│   ├── l2_train_llatisa.json      (Level 2: 45,000+ samples)
│   └── l3_train_llatisa.json      (Level 3: 3,500+ samples)
└── Test/
    ├── l1_test_startend.json
    ├── l1_test_minmax.json
    ├── l1_test_subseries.json
    ├── l1_test_multiseries.json
    ├── l2_test_global.json
    ├── l2_test_local.json
    ├── l2_test_numerical.json
    └── l3_test.json
```

## Task Types

### Level 1 - Basic Time Series Analysis
- Finding maximum/minimum values and their indices
- Trend detection (start/end values)
- Subsequence identification

### Level 2 - Multi-Series Analysis
- Global patterns and relationships
- Local anomalies and features
- Numerical reasoning over multiple series

### Level 3 - Advanced Reasoning
- Complex queries requiring multi-step reasoning
- Real fine-tuning (RFT) and GRPO annotations

## Citation

If you use this dataset in your research, please cite:

```bibtex
@article{llatisa2026,
  title={LLaTiSA: Towards Difficulty-Stratified Time Series Reasoning from Visual Perception to Semantics},
  author={Yueyang Ding, HaoPeng Zhang, Rui Dai, Yi Wang, Tianyu Zong, Kaikui Liu, Xiangxiang Chu},
  journal={arxiv preprint arxiv: 2604.17295},
  year={2026}
}
```

## Licensing

This dataset is released for research purposes.

## Dataset Creator

Created as part of the LLaTiSA project.

## Contact

For issues or questions regarding the dataset, please open an issue on the HuggingFace Hub.