CausalNLP/stride-preproc-climbmix

STRIDE: Preprocessed ClimbMix

Tokenized ClimbMix sequences used for STRIDE's pretraining-data attribution experiments. There is one training pool per released nanochat depth (d12, d16, d20, d24) and one held-out test set shared across depths. These are the corpora the released nanochat operators attribute over, so STRIDE's column indices line up with the lines of these files.

Files

File	Sequences	Size	Contents
climbmix_train_d12.jsonl	1,317,003	3.8 GB	training pool for the d12 base model
climbmix_train_d16.jsonl	2,841,368	8.2 GB	training pool for the d16 base model
climbmix_train_d20.jsonl	5,332,750	15.4 GB	training pool for the d20 base model
climbmix_train_d24.jsonl	8,868,410	25.5 GB	training pool for the d24 base model
climbmix_test_d24.jsonl	500	1.4 MB	held-out test queries, shared across depths

Each depth's training pool is the set of ClimbMix shards that nanochat base model was pretrained on, so larger models have larger pools. The test set comes from shards held out of pretraining, so it is unseen by every base model.

Format

Each line is a JSON object:

{"input_ids": [32759, 1145, ...], "answer_start": 1, "doc_id": 12, "shard": 0}

• input_ids: 512 token ids from the nanochat tokenizer (RustBPE, vocab size 32768). The matching tokenizer ships with rishitdagli/stride-nanochat under tokenizer/.
• answer_start: first token scored by the loss; 1 here, so the whole sequence (next-token prediction) is scored.
• doc_id, shard: provenance of the sequence in the source ClimbMix parquet shards.

How it was built

From the ClimbMix parquet shards with prepare_data (see the code), using the shard range each base model was pretrained on:

python -m stride.cli.prepare_data \
    --shard-dir .cache/nanochat/base_data_climbmix --shard-start 0 --shard-end 127 \
    --output climbmix_train_d24.jsonl --seq-len 512

Usage

In STRIDE the training pool is the attribution corpus (the operator's subsets index into its lines) and the test set holds the queries:

python -m stride.cli.run_influence \
    --operator-dir operators/d12 --base-checkpoint-dir base \
    --test-path climbmix_test_d24.jsonl --output influence_scores.npz

Influence scores are (n_test, n_train): row i is test query i (line i of the test file), column j is training example j (line j of the matching climbmix_train_d*.jsonl).

Related

• rishitdagli/stride-nanochat: base models, trained operators, and the tokenizer.
• rishitdagli/stride-lds: LDS ground-truth losses and the test set.
• Code

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}, 
}