AAMAS2027-Roles / config /SECOND_DOMAIN_SLOT.md
UHXinyang's picture
AAMAS2027-Roles v1.2.0
4790c54
|
Raw History Blame Contribute Delete
3.78 kB

Second domain (SciBench) — SLOT FILLED in v1.1.0

STATUS: FILLED. As of v1.1.0 this release contains SciBench data: 2,296 trajectories, 574 problems, 2 actors, Arms A and B only. This file is kept as the record of what was promised, so the promises can be checked against what was actually done. Each numbered step below is marked with its outcome.

promised done?
arrives as v1.1.0, never an edit to v1.0.0 yes — build and emitter both assert every v1.0.0 file is byte-unchanged
new rows carry benchmark_id='scibench', domain='science', own config_version yes — aed89f28ab338bdb, asserted single-valued over all 2,296 rows
new trace gets its own sha256/rows/bytes; existing three keep theirs yes — raw/MANIFEST_SCIBENCH.json; the v1.0.0 digests are re-listed in VERSION.release_history
CHECKSUMS.txt / VERSION regenerated, history gains an entry, v1.0.0 hashes stay listed yes
per-domain results split in the card yes — Cross-domain (v1.1.0)
SciBench gets its own MIT entry; NOT absorbed into Apache-2.0 yes — plus LICENSES/MIT-SciBench.txt verbatim, verified three ways

One thing this file did not anticipate: the Arm A prompt had to be domain-adapted, so Arm A is not the same contract in the two domains (Arm B is byte-identical). That is stated in the card, in config/run_config_scibench.json, and in a per-prompt diff table in config/prompts_scibench.json. It is why no cross-domain P1a or clause-specific fidelity metric is computed.


The original text, as written before the slot was filled

Status in v1.0.0: NOT PRESENT. Nothing in this release is SciBench data.

A second-domain campaign on SciBench was running while v1.0.0 was packaged. It is deliberately not blocked on: v1.0.0 is complete and self-consistent as a single-domain (competition math, Omni-MATH-2) release.

Why this file exists

So that a later freeze is added, never silently folded into these numbers. Every row in v1.0.0 carries benchmark_id = "omnimath2" and domain = "math" — verified over all 11,600 rows, one distinct value each. A consumer can therefore always separate domains by filtering, including after v1.1.0 lands.

How the second domain will be added

  1. It arrives as v1.1.0, a new minor version — never as an edit to v1.0.0. v1.0.0's files keep their bytes and their sha256s.
  2. New rows carry benchmark_id = "scibench", domain = "science", and their own config_version. Existing rows are untouched.
  3. manifests/FROZEN_MANIFEST.json gains the new trace file with its own sha256/rows/bytes; the three existing entries keep their current hashes.
  4. CHECKSUMS.txt and VERSION are regenerated; VERSION.release_history gains an entry. The v1.0.0 hashes stay listed so anything published against v1.0.0 remains checkable.
  5. The card gains a per-domain results split. Any cross-domain claim is new and must be checked against the paper's claims first — a second domain does not inherit the math findings.

Licensing consequence — do not skip this

SciBench is MIT, not Apache-2.0, and the upstream registry marks it evaluation_only_recommended. It is a different licence from the Omni-MATH-2 material already here. Adding it means:

  • adding a SciBench row to LICENSES/registry.json (its license_id, training_use_policy, and what is actually redistributed), and
  • not extending this release's Apache-2.0 statement over it.

See LICENSES/README.md, which already records the per-source position rather than a single blanket licence, precisely so this slot can be filled without rewriting the licensing story.