Watch-OA Nominal References (v2)
500 frozen ordinary shopping contexts at product-detail decisions, for use with Watch-OA runtime v2.
These retain v1's 500 distinct task families, sampled using seed 42 from
lclan/webshop_expert_trajectories, file webshop_gpt4_0613.json, revision
e13f9187455c57f3f076a3d42c8e06700b018e6c. Known training/reserved families were
excluded. There is zero normalized-family overlap with the 100 runtime families.
Replay each frozen request and ordinary query in pinned WebShop, open the first result regardless of brand, and select only option values uniquely mentioned in the request. Freeze the next-action input. No suspect-model responses, rewards, activation labels, or conflict outcomes are used to select these references. There is no guarantee that a hidden trigger is absent.
The raw observation format matches runtime v2. This conditional candidate-inspection
distribution is different from v1's post-search references; do not mix versions
when calibrating a detector. Provenance is evaluator-only. Only messages from
inputs.jsonl should be supplied as reference input.
- Downloads last month
- 14