aktan-w's picture
Update the card for WSCA v2 (validation set beside v1)
dfbbc54 verified
|
Raw History Blame
3.24 kB
metadata
task_categories:
  - text-generation
language:
  - en
tags:
  - web-search
  - search-api
  - code
  - agents
  - tool-use
  - enterprise-api
  - successfactors
  - ukg
  - veeva-vault
  - genesys
  - sap
  - servicenow
  - workday
size_categories:
  - n<1K
configs:
  - config_name: hard_tickets
    data_files:
      - split: eval
        path: data/eval.jsonl
  - config_name: hard_tickets_v2
    data_files:
      - split: eval
        path: data/eval-v2.jsonl

Web Search for Coding Agents: validation set

322 validation tickets in the v2 format. A coding agent gets a product ticket and the request its code sends today, and must find the API call the ticket needs through one search API.

This dataset is gated and shared with partners only. It is the validation set for the v2 harness. The boards are scored on a held-out private set that is not distributed. Scores computed on these rows are not comparable to the boards.

Files

  • data/eval-v2.jsonl: 322 rows, 1,574,817 bytes, config hard_tickets_v2. Vendors: genesys 29, sap 1, successfactors 101, servicenow 18, ukg 64, veeva-vault 37, workday 72. Every row has set validation, kind patch and implemented true.

The v1 validation file (55 rows) is unchanged and is on the Hub as data/eval.jsonl. It is not copied here.

v2 fields

Rows keep the v1 fields task_id, vendor, ticket, starter_code and gold_code. gold_code is the reference solution. It is used by the offline selftest and is never shown to the agent. The v2 format adds:

field meaning
set validation for these rows
prompt the code-edit prompt (patch mode). The v2 question is built from vendor, the starter's request and ticket
outfile the file the agent edits
kind patch
n_changes number of changes in the reference patch
implemented whether the reference patch is implemented
source_url the vendor documentation page for the answer
docs documentation page URLs for the question. Their hosts, with source_url's host, count as the vendor's documentation in the retrieval check
required tokens the grader looks for in the final call and the search results
query_forbidden gold path tokens; a search query that contains one is logged as a leak
why_opaque why the answer is hard to find by search
retrieval_key answer key for the retrieval check. answers is null on 26 rows, and the retrieval check reports those rows as not evaluated
known_issue set on three rows: the starter's call already passes because the grader ignores version segments, as the locked run did. On each of the three, the starter and the reference solution differ only in the version segment. The selftest exits 0 only when the rows whose starter passes are exactly the rows with this field

Search results in v2

The v2 harness shows search results to the agent under one shared 32,000-character preview budget. v1 capped each hit at 1,200 characters. Runs under the two settings are not directly comparable.

Harness

The harness, the scoring and the vendor runners are in openbenchmarks-labs/web-search-for-coding-agents.