SequentialDataVal

Official implementation for Unifying and Optimizing Data Values for Selection via Sequential Decision-Making. Accepted to ICML 2026 as a Spotlight Paper.

This repository contains the algorithms, configs, runners, and plotting/comparison scripts used for the paper experiments. Large datasets, model checkpoints, generated embeddings, and full run outputs are distributed separately when needed.

Installation

Use Python 3.9 for the OpenDataVal experiments. The paper code was tested with OpenDataVal 1.3.0, which is not compatible with Python 3.12.

pip install -r requirements.txt

Or with conda:

conda env create -f environment.yml
conda activate sequentialdataval

For development:

pip install -e ".[all]"

Tested core stack: Python 3.9, OpenDataVal 1.3.0, NumPy 1.25, pandas 1.5, scikit-learn 1.2+, PyTorch 2.1.2. Utility approximation additionally needs cvxpy; DATE-LM fine-tuning additionally needs the GPU dependencies listed in finetuning/requirements-gpu.txt.

Quick Check

bash scripts/smoke_test.sh
python src/main.py --dataset 2dplanes --seed 10 --config config/base_config.yaml

Paper Experiment Entry Points

RQ1 exact dynamic-programming gap and detailed RQ1 Figure 6:

python src/main.py --dataset digits --seed 10 --config config/rq1_config.yaml \
  --skip_bipartite --include_dp --dp_max_subset_size 20

bash scripts/run_rq1_dp.sh
python scripts/aggregate_rq1_dp.py
python scripts/plot_rq1_figure6.py \
  --results_root results_rq1_dp \
  --out_dir results/rq1_figure6

RQ2 curvature:

python scripts/run_rq2_curvature.py

RQ3 selection curves:

bash scripts/run_rq3.sh

Per-seed outputs are written under results/<dataset>/seed_<seed>/addition_experiment_results.csv. The full RQ3 script runs 8 datasets x 20 seeds and is CPU-heavy; use scripts/smoke_test.sh first, or edit the datasets/seeds arrays in scripts/run_rq3.sh for a short verification run.

Budget-500 table:

python scripts/run_budget500_curve_mean_table.py \
  --curve_step 10 \
  --max_workers 6 \
  --out_dir results/selection_results_500_curve_mean

python scripts/compare_budget500_table3.py \
  --summary results/selection_results_500_curve_mean/summary_curve_mean.csv

Budget-500 uses train_count=500; standard datasets use valid/test=50/500, while digits and bbc-embeddings use valid/test=100/1000. The reported value is the mean over the top-k selection curve, not only the endpoint at k=500. Use --max_workers 1 for bbc-embeddings on memory-constrained machines.

Utility approximation:

cd utility_appro
pip install "cvxpy>=1.4.0"
bash run_experiments.sh

DATE-LM Fine-Tuning Extension

The finetuning/ directory contains the BipCov selection code and DATE-LM-style orchestration scripts. Running these experiments requires external DATE-LM data, model checkpoints, and benchmark files.

Prepare the following inputs before launching the fine-tuning pipeline:

• DATE-LM checkout
• Tulu3 instruction pool
• Llama-3.1-8B checkpoint in the format expected by DATE-LM/LitGPT
• MMLU, GSM8K, and BBH evaluation data

Typical setup on a GPU machine:

pip install -r finetuning/requirements-gpu.txt
export SEQDATAVAL_ROOT=/path/to/SeqDataVal
export DATELM_ROOT=/path/to/DATE-LM

bash "$SEQDATAVAL_ROOT/finetuning/scripts/datelm_paper/patch_datelm_table3.sh" \
  "$DATELM_ROOT"

Multi-seed train/eval robustness:

python "$SEQDATAVAL_ROOT/finetuning/scripts/datelm_paper/run_multiseed_train_eval_only.py" \
  --datelm_root "$DATELM_ROOT" \
  --python python \
  --seeds 42 1337 2025 \
  --methods random rds_plus bipcov \
  --tasks mmlu gsm8k bbh \
  --mmlu_eval_batch_size 32 \
  --gsm_eval_batch_size 16 \
  --bbh_eval_batch_size 16 \
  --skip_verify \
  --cleanup_checkpoints

BipCov from precomputed embeddings:

python finetuning/bipcov/probe_bipcov_from_emb.py \
  --train_emb /path/to/train_emb.npy \
  --ref_emb /path/to/ref_emb.npy \
  --out /path/to/scores/bipcov_metrics.npy \
  --k_max 10000

DATE-LM protocol used in the paper: 200k Tulu3 pool, 10k selected examples, 100 prompt+label reference examples per task, Llama-3.1-8B base model, LoRA rank 128 / alpha 512 / dropout 0.1 on q/k/v/o modules, learning rate 2e-5, effective batch size 128, 2 epochs, max sequence length 2048, bf16 precision, and train seeds 42, 1337, 2025.

Repository Layout

src/                         Core algorithms and OpenDataVal integration
config/                      Paper experiment configurations
scripts/                     CPU experiment, comparison, and plotting runners
utility_appro/               Utility approximation experiments
finetuning/                  DATE-LM/BipCov selection and fine-tuning scripts
requirements.txt             Python dependencies
environment.yml              Conda environment

OpenDataVal downloads/preprocesses datasets relative to the working directory. For repeated runs, create a persistent repo-root cache:

mkdir -p data_files

Datasets And Methods

OpenML datasets: 2dplanes, nomao, bbc-embeddings, MiniBooNE, digits, election, electricity, and fried.

Core methods:

• BipartiteMatchingEvaluator: bipartite graph approximation
• DynamicProgrammingEvaluator: exact optimal sequential selection

OpenDataVal baselines:

• Random, Leave-One-Out, InfluenceSubsample
• Data Shapley, Beta Shapley, Data Banzhaf
• AME, DVRL, Data-OOB

Citation

@article{chi2025unifying,
  title={Unifying and Optimizing Data Values for Selection via Sequential-Decision-Making},
  author={Chi, Hongliang and Wu, Qiong and Zhou, Zhengyi and Light, Jonathan and Dodwell, Emily and Ma, Yao},
  journal={arXiv preprint arXiv:2502.04554},
  year={2025}
}

License

MIT. See LICENSE.