UQE

UQE usage

UQE is a standalone implementation in this checkout; it is not registered in the root meta-orchestrator. Its paths are relative to the current working directory, so run it from systems/UQE.

Run

cd systems/UQE

# Show the CLI
python main.py --help

# Run one dataset and override its default query type
python main.py --dataset lcr --query-type SFW

# Run every registered dataset
python main.py --all

Supported --dataset values are art, art_image, disease, drug, finance, institutes, lcr and player. --query-type is passed unchanged to the selected dataset handler; typical folders are named SF, SFW, or another workload-specific code.

Required local data and queries

The repository does not currently version the systems/UQE/data/ and systems/UQE/query/ payloads expected by these runners. Supply them before execution:

systems/UQE/
|-- data/<dataset>/dataset.json
`-- query/<dataset>/<query-type>/*.sql

Dataset directory casing/names follow the handlers in main_*.py (for example data/Finance with query/Finance, and data/player with query/player). A dataset JSON record can contain text directly and paths for multimodal inputs, for example:

{
  "id": "1",
  "description": "your text",
  "image": "path/to/image"
}

Configuration

config_uqe.py defines the runtime knobs:

• USE_BART, BATCH_SIZE, BUDGET
• AGGR_STRATEGY, N_CENTROIDS, N_ITER
• GROUP_EXTRACT_SAMPLE_RATIO, AGGR_CLUSTER_SAMPLE_RATIO
• MODEL, OPENAI_KEY, BASE_URL

This folder does not ship a requirements file. The imports require, among other packages, OpenAI's client, pandas, PyTorch/torchvision, Transformers, Sentence Transformers, scikit-learn, sqlparse and tqdm.

Outputs

Each handler writes beneath systems/UQE/result/<dataset>/<query-type>/<timestamp>/<query-name>/<query-name>.csv. The exact dataset component is defined by the handler and may preserve casing (for example Finance).