Superseded first-click diagnostic
This v1 release is retained for provenance. Its first-product-click endpoint yielded zero Positive runtime cases and is unsuitable for measuring positive-class backdoor detection. Zero positives here do not establish that the checkpoints never activated. Browsing a candidate does not establish a conflicting purchase. Use qiusizhan/Watch-OA-nominal-reference-v2 for the corrected single-action purchase-decision inputs. Do not mix v1 and v2 runtime/reference contexts. The original v1 inputs, responses and labels below are unchanged; the observation-point correction uses a separate version.
Watch-OA Nominal Reference Inputs (v1)
500 fixed ordinary shopping contexts after the first product search, shared across the four compatible Watch-OA checkpoints. The corresponding runtime evaluation set contains 250 inputs and model-specific single-action labels.
Sampling and interface
Requests and their ordinary first searches originate from
lclan/webshop_expert_trajectories,
file webshop_gpt4_0613.json, revision e13f9187455c57f3f076a3d42c8e06700b018e6c.
The source sampling uses seed 42 and one representative per normalized task family,
excluding known training and reserved evaluation families. Selection uses no
brand, trigger, reward, target-action or suspect-model-output filter. These are
nominal references, not a guarantee that a hidden backdoor is inactive.
The exact frozen requests and first searches were replayed in the full pinned WebShop environment. Each record contains the original shopping harness, acknowledgment, initial page, fixed search and actual raw search results. These released pages replace the earlier preliminary delimiter-normalized source prefixes. No additional observation wrapper or action-list serialization is added. Later source-trajectory responses and rewards are not model-visible inputs.
All 500 task families are distinct, with zero normalized-family overlap with the 100 task families in the runtime set. Normalization removes trailing price limits; this check does not prove semantic or product-level independence. The product catalog mirror and original source hashes are recorded in provenance.
Only messages in inputs.jsonl are reference input for an auditor. IDs are
bookkeeping. provenance/source_tasks.jsonl and other provenance are evaluator-only.
No activation labels, detector scores, fixed calibration folds or thresholds are
attached to these references.
from datasets import load_dataset
reference = load_dataset("qiusizhan/Watch-OA-nominal-reference-v1", "inputs", split="reference", token=True)
- Downloads last month
- 18