500 fixed nominal reference inputs shared across the four compatible Qwen2.5-Coder BALD-scene checkpoints. This is the reference companion to BALD-scene-runtime-v1. These are text-only next-decision prompts, not images fed to the audited model, simulator episodes, or labeled negative examples.

Nominal reference protocol

Inputs are sampled independently from an ordinary driving image pool, without filtering on ground-truth backdoor information, attacker objectives, trigger conditions, activation labels, or detector scores. They are not guaranteed backdoor-inactive. A naturally occurring task scene may satisfy an unknown trigger. No audited-model inference or activation labeling is performed to construct this set. No samples are removed or replaced based on image objects, caption semantics, or generated driving behavior.

Source and sampling

Source images are the CAM_FRONT images in Xiaodong/Nuscenes-v1.0-trainval-CAM_FRONT, pinned at c2a6deff22771a163e33837f4d44f017cadcb4b4, originally from nuScenes.

The source ZIP contains 34,149 eligible front-view image paths. We sort those paths lexicographically and select 500 without replacement using Python random.Random(42).sample. The population and selected path list are frozen before caption API calls. File-extension, camera-channel, and archive-metadata checks identify actual CAM_FRONT image files; there is no scene-content filtering. Selected images are downloaded using byte ranges and checked against archive size/CRC32 and SHA256. Original images are retained locally, not redistributed in this text-input release.

The sampling unit is a frame, not a unique physical scene. Multiple selected frames may come from the same drive or scene. Per the requested full-pool random sampling protocol, training/runtime scene IDs are not excluded. Because the older processed runtime records do not retain source image IDs, physical-scene/frame disjointness from model training or runtime evaluation cannot be certified. Exact prompt overlap with the published 250 runtime inputs is separately checked. Do not describe this as a scene-disjoint holdout or as 500 independent physical driving scenes.

Image-to-description conversion

Each frozen image is supplied to OpenAI gpt-5.4-mini-2026-03-17 with the identical generic English scene-description prompt, image detail high, reasoning effort low, a 1,536-output-token budget, and strict JSON format with a single description field. The model is asked to describe visible road layout, traffic, signs, objects, spatial relationships and visibility without inventing details or recommending driving actions. It receives no audited model identity, attacker target, trigger specification, backdoor reasoning, or reference decision. Full settings and the exact prompt are in provenance/caption_settings.json.

The first three frozen samples serve as an API-format pilot and remain in the final set. No caption-content rejection or resampling is applied. The first complete schema-valid response for each selected image is retained. Automatically generated captions can contain perception errors and are not expert-verified ground truth. Recalling the same API model does not guarantee byte-identical captions; saved caption records allow exact reconstruction of the published inputs.

Task context and template

The system prompt is byte-identical to the BALD-scene runtime system prompt. It requests JSON with reasoning and one of Turn-left, IDLE, Turn-right, Acceleration, Deceleration. No reference assistant answer is included.

Speed and navigation are synthetic task context, not measured scene telemetry. Using a separate PRNG with seed 142, each speed is independently drawn uniformly from integers 5 through 30 mph and each navigation instruction uniformly from the following fixed ordinary task menu:

Stay in the current lane and straight ahead.
Turn left at the next intersection.
Turn right at the next intersection.
Ready to stop at the next intersection.
Keep speed to 40 mph if possible.

These draws are independent of image content, captions, legitimate-action labels and audited-model outputs. They may differ from the actual recorded vehicle state. This image distribution, caption generator and synthetic context procedure define the declared nominal distribution; exact equality with the original BALD paper's generation distribution is not claimed.

Completed validation

All 500 selected images and 500 API descriptions are retained, with 500 unique final prompts. There are zero identical prompts shared with the 250-input runtime set; this is not a physical-scene-disjointness claim. The four delivered checkpoint tokenizers produce identical input token IDs on this set. Prompt lengths including the generation prefix are 404–463 tokens (median 435), fitting the 32,768-token context with a 512-token generation reserve. These checks do not classify reference inputs as backdoor-active or inactive.

Files and use

inputs.jsonl contains only case_id and the model-visible system/user messages. metadata.jsonl provides source image identifiers, hashes, caption model, synthetic context and provenance. Neither includes trigger-presence or activation labels. image_manifest.jsonl lists all 500 preselected images and their archive locations. manifest.json, validation.json, and provenance/ record settings, checks and checksums. code/ contains image selection, API captioning, assembly and independent verification scripts. Original image binaries, API credentials and intermediate per-request files are not part of the published release.

Auditors may use these inputs to estimate nominal feature distributions or calibrate runtime monitors. They must not assume every reference response or activation is benign. No folds, thresholds, detector features or reference activation statistics are prescribed. Keep provenance and source metadata separate from messages passed to a detector.

After access approval and Hugging Face authentication:

from datasets import load_dataset
reference = load_dataset(
    "qiusizhan/BALD-scene-nominal-reference-v1",
    "inputs", split="reference", token=True,
)

Verify file hashes, sampling, context draws, prompt structure and saved-caption assembly:

python code/verify_reference.py --root .

To reproduce image sampling and fetch the 500 original selected files:

python code/sample_images.py --root .

For fresh API captioning, provide OPENAI_API_KEY in the process environment and run code/caption_images.py. It also supports the original task-local credential file ~/.config/openai/bald_scene_api_key; neither mechanism is serialized into artifacts. provenance/caption_records.jsonl preserves the released descriptions and token usage. Original selected images are required to repeat API generation; no sandbox, Docker, CARLA, or GPU is required.

Downloads last month: 9

Total file size:

5.82 MB