strifero/attribution-groundtruth

Ground-truth training-data attribution

A benchmark for methods that claim to tell you which training data taught a model something. Two small models were trained from scratch on the same ordered 5B-token stream, differing only in which half of 528 fictional facts their planted documents stated. Because we planted the documents, we know the right answer; because we have the twin that never saw each fact, we can measure what each document actually did.

• Models: tda-124m-A and tda-124m-B, with checkpoints every 0.5B tokens
• Report: Paper trail
• Code (stream builder, training, every method, scoring): https://github.com/strifero/attribution-groundtruth

All planted content is fictional. The base training text is not redistributed here; stream/ holds the exact recipe and order to rebuild it, with a hash of every window to check the rebuild.

Files

Path	What it is
facts/spec.json	The 528 facts: entity, fact, answer, cloze prompt, 20 candidate answers, frequency (2, 8, 32 or 128), position (early 0 to 20%, middle 40 to 60%, late 85 to 100% of training), run (A or B), and the distractor topics
documents/plant_docs.jsonl	22,440 planted documents and 1,050 distractors (two per entity, saying nothing about the planted fact). row is the id used everywhere else
stream/index_{A,B}.npz	One row per training window, in training order (4,882,688 windows of 1,025 tokens; 256 windows per step). src, shard, off: where a base window came from (source id in the order python_edu, python_stack, fineweb_edu, cosmopedia; shard file; token offset). plant_row: the planted or distractor document in that window, or -1. window_blake2b64: BLAKE2b, 8-byte digest of the window's uint16 tokens, read as little-endian uint64
stream/stream_meta.json	Mix, window counts per source and the shard files used
eval/{A,B}/queries.jsonl	Two queries per fact (cloze and question form) with the answer and relevant: positions in the pool of that fact's planted windows
eval/{A,B}/pool.npz	The 112,270 candidate windows a method ranks: every planted and distractor window plus 100,000 random base windows. window is the stream index, kind is planted, distractor or base
results/{A,B}/fact_eval_final.json	Per fact, under that run's final model: rank of the true answer among the 20 candidates, margin over the best wrong candidate, and logp_true, for both prompt forms. The causal effect of a run A fact is its logp_true in A minus its logp_true in B
results/{A,B}/res_*.json, mass_*.json	Our per-query retrieval scores (AP, recall@10, R-precision) and attribution mass for each method and checkpoint
results/causal*_final.json	Correlation of each method's attribution mass with the causal effect, raw, within design cell, and partial (controlling answer length and the unexposed model's loss), with bootstrap 95% intervals

How to grade a method

1. For each fact the model learned, compute the causal effect: log p(answer | model that saw it) minus log p(answer | twin). results/*/fact_eval_final.json has what you need to pick the learned facts.
2. Score every pool window for each query with your method.
3. Retrieval: average precision of the fact's planted windows among the 112,270.
4. Causal agreement: the mean z-score of the planted windows (attributed mass), correlated with the causal effect across facts, ideally within design cells and controlling answer length and the twin's loss. attr/causal3.py in the code repository does this.

Our results

Method	Retrieval AP	Partial r with causal effect
BM25	0.995	-0.02
TracIn, preconditioned	0.82	+0.04
Influence functions (Gauss-Newton)	0.74	+0.03
bge-base embeddings	0.61	+0.15
TracIn, dot	0.15	+0.06
Gradient cosine	0.08	+0.09
Gradient dot	0.05	+0.09
Hidden-state cosine	0.003	+0.13

Finding the documents is easy here (the fictional names are unique), so the retrieval column is a sanity check. The causal column is the hard part, and no method we tried does well on it. The report has the details.

Rebuilding the training stream

You need the four source corpora, Elroy's tokenizer and shard builder (github.com/strifero/elroy, scripts/fetch_stack_edu.py and scripts/build_shards.py), and dl.sh, fetch2.sh, pipeline.sh and build_stream.py from the code repository. build_stream.py is deterministic: the same shards give the same two streams, and the window hashes in stream/index_*.npz confirm it.

Licenses

Our files are Apache-2.0. The base corpora keep their own terms: Stack-Edu (HuggingFaceTB, Python subset; files with a permissive or no detected license, as in The Stack v2), StarCoderData (bigcode, permissive), FineWeb-Edu (ODC-By), Cosmopedia-v2 (ODC-By). The planted and distractor documents were written by Qwen3.5-9B (Apache-2.0) from fictional facts, and are released in the dataset repository.