# Format

There is no single top-level `annotations.jsonl` in this directory. Instead, the annotations are split across these 5 subdirectories:

- `clean_TSQA/anomaly_detection/annotations.jsonl`
- `clean_TSQA/classification/annotations.jsonl`
- `clean_TSQA/open_ended_qa/annotations.jsonl`
- `clean_cats-bench_hard/annotations.jsonl`
- `clean_timeomni/annotations.jsonl`

Each line is one JSON object (JSONL format). The core fields are described below.

## Common Fields

### `id` (string)
Unique sample ID.

- TSQA examples: `ts_anomaly_0`, `ts_classif_1`, `ts_openqa_3`
- TimeOmni example: `1_scenario_understanding_test`
- CATS hard example: `ts_retrieval_perturbed__agriculture_100_test__0`

This field is generated by preprocessing scripts and is used for deduplication, tracking, and evaluation alignment.

### `image` (string)
Relative path to the image file (relative to the current sub-dataset directory).

- TSQA / TimeOmni: usually `images/*.png`
- CATS hard: usually `plots/*.jpeg` (can also be `.jpg/.png/.webp`)

During training/inference, the model should load this image before answering the question.

### `answer_type` (string)
Answer type, which determines how outputs should be evaluated.

- `mcq`: multiple-choice (answer is typically an option letter such as `A/B/C/D`)
- `exact`: exact text match (e.g., `Yes/No`, `True/False`, numbers)
- `approximation`: approximate numeric answer (supported by source scripts; rarely seen in this `hg_dataset` snapshot)

Source scripts:
- TSQA: `tsqa.py`
- TimeOmni: `testomni.py` (always `mcq`)
- CATS: `cats.py` / `cats_test.py`

### `conversations` (array[string, string])
Two-element array:

- `conversations[0]`: question/prompt
- `conversations[1]`: gold answer

This dataset uses a simplified two-turn format, not a role-tagged format like `{from: human/gpt}`.

## Extra Field in CATS hard

### `task_type` (string, only in `clean_cats-bench_hard`)
Task subtype. This field exists only in CATS hard annotations and distinguishes retrieval task variants.

Common value in the current data:
- `ts_retrieval_perturbed`

Additional possible values supported by `cats_test.py`:
- `ts_retrieval_cross_domain`
- `ts_retrieval_same_domain`
- `caption_retrieval_cross_domain`
- `caption_retrieval_perturbed`
- `caption_retrieval_same_domain`

## Minimal Examples

```json
{"id":"ts_anomaly_0","image":"images/ts_anomaly_0.png","answer_type":"exact","conversations":["...question...","No"]}
```

```json
{"id":"ts_retrieval_perturbed__agriculture_100_test__0","image":"plots/agriculture_100_test.jpeg","answer_type":"mcq","task_type":"ts_retrieval_perturbed","conversations":["...question...","A"]}
```

## Mapping to Preprocessing Scripts

- TSQA subsets: `/home/xinyu/ChartModel/chart/app/data_process/rl/tsqa.py`
- CATS hard: `/home/xinyu/ChartModel/chart/app/data_process/rl/cats_test.py`
- TimeOmni: `/home/xinyu/ChartModel/chart/app/data_process/rl/testomni.py`

These scripts define field generation logic, `answer_type` assignment, and question text cleaning rules.

## Inference Example

Below is an example workflow using vLLM OpenAI-compatible server plus `bon_filter.py`.

### 1) Start vLLM server

```bash
CUDA_VISIBLE_DEVICES=4,5,6,7 python -m vllm.entrypoints.openai.api_server \
  --model /path/Qwen3-VL-2B-Instruct \
  --host :: \
  --port 8003 \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.85 \
  --limit-mm-per-prompt '{"image": 1}' \
  --data-parallel-size 4 \
  --trust-remote-code \
  --max-num-batched-tokens 8192
```

### 2) Run inference with `bon_filter.py`

Example on TimeOmni subset:

```bash
python chart/app/rl/bon_filter.py \
  --image_dir /hg_dataset/clean_timeomni \
  --input_jsonl /hg_dataset/clean_timeomni/annotations.jsonl \
  --output_jsonl /hg_dataset/clean_timeomni/bon.jsonl \
  --model_name /path/Qwen3-VL-2B-Instruct \
  --n 1
```

You can switch `--image_dir` and `--input_jsonl` to other subsets in this dataset, for example:

- `/hg_dataset/clean_TSQA/anomaly_detection`
- `/hg_dataset/clean_TSQA/classification`
- `/hg_dataset/clean_TSQA/open_ended_qa`
- `/hg_dataset/clean_cats-bench_hard`

Note: ensure the `image` field inside each JSONL is a valid relative path under `--image_dir`.