CausalNLP/stride-lds

STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbations

Ground-truth Linear Datamodeling Score (LDS) targets for the four nanochat pre-training models, plus the shared held-out test set.

Each lds_<tag>.jsonl was produced by: sampling a pool of pre-training examples, drawing 256 random 30%-subsets, training a fresh nanochat from scratch on each subset, and recording per-example held-out test losses. The _meta header records the pool indices so scores defined over the full corpus can be compared to the subset losses.

Files

climbmix_test_d24.jsonl    # 500 held-out test queries (shared across depths)
lds_d12.jsonl              # 256 subsets x 500 test losses  (+ lds_d12.meta.json)
lds_d16.jsonl
lds_d20.jsonl
lds_d24.jsonl

Each non-_meta line:

{"train_subset": [pool-local indices], "train_subset_global": [corpus indices], "test_loss": [500 floats]}

Usage

from stride.inference import Stride
attr = Stride.from_pretrained("d12")
result = attr.attribute(test_queries)
lds = attr.evaluate_lds(scores=result.scores)   # downloads lds_d12.jsonl from here
print(lds["lds_spearman_mean"])

Or score an existing method's .npz:

python -m stride.cli.eval_lds --scores my_scores.npz --lds lds_d12.jsonl --method MyMethod

The models and operators live in the companion model repo rishitdagli/stride-nanochat.

Citation

@misc{dagli2026stridetrainingdataattribution,
      title={STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbations}, 
      author={Rishit Dagli and Abir Harrasse and Luke Zhang and Florent Draye and Amirali Abdullah and Bernhard Schölkopf and Zhijing Jin},
      year={2026},
      eprint={2606.05165},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2606.05165}, 
}