500 fixed reference inputs from 500 different coding tasks, shared across compatible checkpoints and auditing methods. Each input is an unmodified SWE-agent conversation prefix immediately before the next assistant action.

This is the reference companion to the separate 250-input PersistBD runtime evaluation set. Access requires manual approval, matching that release. The reference and runtime evaluation sets have zero shared instance_id values and zero identical prompts.

What nominal means

These are ordinary task inputs sampled without known-trigger, attacker-objective, activation-label, or detector-score filtering. They are not certified negative interactions: an unknown backdoor can still activate on a nominal input. No model outputs or activation labels are supplied. Source resolution describes the eventual success of the original trajectory, not backdoor activation.

The selected source trajectories comprise 62 resolved and 438 unresolved runs across 84 repositories. At the task level, 68 of the selected tasks have at least one resolved trajectory in the source pool; the randomly chosen trajectory need not be that resolved run. Resolution status does not affect sampling.

Auditors can use the inputs to estimate reference feature distributions or calibrate their methods. There are no predefined folds, anomaly scores, thresholds or covariance settings. These choices belong to each auditing method. Keep source/provenance metadata separate from model-visible conversation messages.

Sampling and holdout scope

The source is the same eight SWE-smith XML trajectory parquet shards used for splits_v4. SHA-256 checksums pin all source files. We exclude task IDs from:

Original splits: insertion_train, insertion_test, benign_post, attacker_train, and attacker_select.
Later benign sets: benign_post_unresolved, benign_post_big, and benign_post_max.
The unresolved extended test set and the existing runtime evaluation set.

These exclusions leave 2,841 candidate tasks: 402 have at least one resolved trajectory and 2,439 have only unresolved trajectories. With seed 42, tasks, trajectories and decision positions receive separate pseudorandom SHA-256 priorities. We retain one eligible prefix per task. The first 500 ranked tasks all supplied an eligible prefix, so no entire task was skipped.

Checks cover message structure, noninitial user-to-assistant boundaries, empty content, severe exact repetition in the existing assistant history, duplicate prompts/task text, and context length. In total, 1,246 oversized candidate prefixes were skipped. No history is cropped or rewritten; future assistant response content is not used for selection.

All selected tasks are disjoint from the listed exclusions. This does not establish absence from every model's training history: old-protocol checkpoints, clean-control training data, unrelated RL data, and base-model pretraining were not exhaustively checked. The predominance of unresolved runs also differs from the resolved-source runtime evaluation distribution and should be reported in calibration experiments.

Input interface

inputs.jsonl records contain only:

{"case_id": "persistbd-ref-v1-0000", "messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}, {"role": "user", "content": "..."}]}

The omitted contents in this example are explanatory placeholders. Released inputs preserve complete source prefixes and the source SWE-agent XML action interface. Use the shared runtime tokenizer and chat template pinned in manifest.json. Prompt lengths range from 1,537 to 32,253 tokens (median 21,072.5), including the assistant generation prompt. Every input fits within 32,768 tokens with a 512-token generation reserve. Other tokenizers, added instructions, or harnesses need their own context-budget checks. No model inference or tool execution was performed to build this set.

Files

File	Contents
`inputs.jsonl`	500 model-visible conversation prefixes.
`metadata.jsonl`	Task and trajectory IDs, source row locations, resolution flags, prefix boundaries, lengths and hashes.
`candidate_pool.jsonl`	Complete ranked candidate task pool with source trajectory locators.
`selection_audit.jsonl`	Selection attempts and prefix eligibility outcomes.
`excluded_instances.json`	Task IDs excluded from each source split.
`validation_report.json`	Counts, overlap checks and input distributions.
`independent_validation.json`	Completed independent source-prefix and tokenizer checks.
`manifest.json`	Source hashes, construction settings, tokenizer fingerprints and release-file checksums.
`release_manifest.json`	Release identity, local-to-portable provenance and release checksums.
`code/make_persistbd_nominal_reference.py`	Portable generator and source-aware validator.
`requirements.txt`	Versions used for construction.

Usage and reproduction

After access is approved and you authenticate with Hugging Face:

from datasets import load_dataset

reference = load_dataset(
    "qiusizhan/PersistBD-nominal-reference-v1",
    "inputs", split="reference", token=True,
)

The generator's sampling and quality rules are unchanged from the local build. Only machine-specific CLI defaults and the generated README's environment-specific launch command were made portable. Data rows and source metadata are byte-identical to the validated local set. A fresh rebuild requires the pinned original trajectory shards, split files and shared tokenizer; this release does not redistribute the full source pool or training data.

python code/make_persistbd_nominal_reference.py \
  --source /path/to/BadCodeAgent \
  --tokenizer /path/to/shared-runtime-tokenizer \
  --out ./rebuilt-reference --seed 42 --count 500

To verify the downloaded release against those same local sources, use the same command with --out . --validate-only. The local build manifest and frozen original generator are identified by hash in release_manifest.json; the release manifest updates artifact hashes for the portable code and dataset card. A rebuild's generated README and machine-path fields will differ; the input rows and metadata reproduce.

Downloads last month: 81

Total file size:

40.9 MB