Where the evaluation corpus came from
This page answers one question: for every prompt in the bundled evaluation track, where did it come from and what is known about it?
This page is about origins. The evaluation track card is about limitations. If you want to know what is wrong with the corpus, read the card, which lists every known defect. If you want to know where a row came from, read this. Neither page repeats the other, deliberately: two documents describing one corpus is how a corpus ends up with two different official stories.
Why this page exists
Four of the five known defects in this corpus were found by somebody reading it, not by anything checking it. That is a bad way to find out that your ruler is bent. So the corpus now carries its own provenance record, and every check described here is something a reader can run.
The three sources
| Source | Side | Rows it contributes | Declared licence | Revision read |
|---|---|---|---|---|
Bahushruth/abliteration-harmful-enriched | harmful | most of the harmful side | apache-2.0, in cardData and tagged | f29c0b77 |
mlabonne/harmful_behaviors, reached through the above | harmful | 176, measured 2026-10-01 | none declared | 01cead01 |
mlabonne/harmless_alpaca | harmless | the harmless side | none declared | 02c6a92c |
Two of the three declare no licence at all. The card explains what we believe the true position is for each, and why that is a belief rather than a fact. The short version: the 520 rows in mlabonne/harmful_behaviors are, by row count and by inspection, AdvBench's harmful_behaviors.csv from the llm-attacks repository, which is MIT licensed. That is an inference from the contents, and neither dataset card says so.
Of those 520 upstream rows, 176 rows of this track's harmful side carry one of their requests, measured template-aware on 2026-10-01. An earlier estimate of "about 430" circulated in this project's docs and was not supported by measurement; it has been corrected everywhere it appeared.
We distribute the track under CC BY-NC 4.0, the most restrictive licence in the chain.
How the rows are laid out, and why that matters
The track holds three datasets, and they are not three partitions. Three partitions are packed into them:
bad_ds harmful fit
bad_eval_ds harmful search ++ measure (in that order)
good_ds harmless fit ++ search ++ measure (in that order)So the boundaries between partitions are recorded in the manifest, not derivable from the files. A consumer that reads the first N rows of good_ds is reading across the fit boundary once N exceeds the fit count, and nothing about the file says so.
That is a real hazard and the tool guards against it. senbonzakura.track.flag_violations reads the manifest and refuses any combination of flags that would cross a recorded boundary. It runs on the abliteration path, on the head-to-head stager, and on senbonzakura validate.
A track with no manifest is unguarded. The check needs recorded boundaries to check against, so a track built before manifests existed skips it silently. If you hold such a track, rebuild it or audit it:
python -m senbonzakura.track --out <track> --auditThe counts
| Dataset | Rows |
|---|---|
bad_ds | 259 |
bad_eval_ds | 4,636 |
good_ds | 4,982 |
Two figures worth knowing if you are configuring a run, measured 2026-09-11: the harmless fit + search partitions hold 385 rows between them and the harmful ones hold 391. Some other tools default to asking for 400 prompts a side, which is more than this track keeps aside for fitting and selection, so a run at that default would be refused rather than quietly reading into the measurement rows.
What is not recorded, and the size of the gap
Part of the harmless side cannot be reproduced. Rows were added to the harmless side after the first build and the addition was not recorded: not the source, not the count, not the revision. So senbonzakura track build reproduces a corpus of the same shape from the same upstreams, and it does not reproduce these rows.
The gap was measured on 2026-10-01 and it is large. Of the 4,982 rows in the shipped good_ds, 2,036 (40.87%) have no request present in mlabonne/harmless_alpaca, the only harmless upstream this track declares. They are spread through all three partitions rather than appended as a tail:
| Partition | Rows with no recorded source | Of |
|---|---|---|
fit | 106 | 257 |
search | 58 | 128 |
measure | 1,872 | 4,597 |
| total | 2,036 | 4,982 |
Those 2,036 rows are 386 distinct requests, not 2,036 independent ones. Each appears about five times under a different template. The 2,946 rows that do trace to harmless_alpaca are 2,946 distinct requests, one row each, with no template expansion. So the unreproducible part of the harmless side is a second source of 386 requests that was template-expanded and then mixed in before partitioning, which is why it is distributed evenly rather than sitting at the end.
Two things follow, and the second is the one that matters for anybody planning a rebuild. The gap to close is 386 requests rather than 2,036 rows, which is a tractable recovery job rather than a hopeless one. And the harmless side is 41% template-expanded against a harmful side that is expanded throughout, so the two sides of this track do not have the same internal structure. Treat the harmless side as partly unreproducible and do not assume a rebuild will give you the same corpus.
How to reproduce this measurement: read good_ds and mlabonne/harmless_alpaca, key both with senbonzakura.track.request_key over templates discovered across the union of the two, and count good_ds rows whose key is absent upstream. Matching by raw string instead gives a different and wrong answer, for the reason the contamination checker documents.
What the count does not cover, stated so it does not read as cleared. It tests against the one harmless upstream this track declares. If those 386 requests came from an undeclared second dataset rather than being written by hand, they have a source that was simply never recorded, which is a recoverable situation rather than a permanent one, and this count cannot tell the two apart. wangzhang/abliterix-datasets is the obvious candidate to check next. It was not checked and not cleared: the copy available when this was measured held only a ref with no data, so there was nothing to compare against.
Every track built after that point carries a manifest recording its sources and revisions, so this cannot happen again. It did happen once, and it happened to the corpus behind the published figures.
Checking the corpus against a public benchmark
AdvBench is inside this track, arriving through mlabonne/harmful_behaviors. If any of those rows sit in the fit or search partitions, then a model tuned with this track was tuned on requests AdvBench would later mark it against, and any AdvBench score from it is in-sample rather than held out.
That check has now been run, and the answer is that this track is contaminated. Measured 2026-10-01: of AdvBench's 508 distinct requests, 8 are in fit, 3 are in search, 159 are in measure and 338 are absent. So 11 requests were fitted or searched on, and an AdvBench figure over the whole benchmark from this track is in-sample. A clean figure is still available over the 497 requests that are either in measure or were never held, and a figure reported that way has to say so in those words. The card carries the table and the caveats on how it was measured.
The same applies to any other public set assembled from these upstreams, which is most of them. Check before you rely on a figure:
python -m senbonzakura.track --out <track> --contamination advbench.txt \
--contamination-name AdvBenchAdd --fail-on-contamination to make it an error rather than a report, and --contamination-report <path> to keep the result as an artefact.
Citing
Cite AdvBench (Zou, Wang, Carlini, Nasr, Kolter and Fredrikson, Universal and Transferable Adversarial Attacks on Aligned Language Models, 2023, MIT licence), Alpaca, the two HuggingFace datasets the builder fetches, and this repository. If you publish a number from the harmful side, publish the harmless arm beside it.