license: mit
language:
- en
tags:
- rlvr
- reinforcement-learning
- alignment
- reasoning
- verifiable-rewards
pretty_name: AmberTrace — Certified Decision Eval Suites
AmberTrace — Certified Decision Eval Suites
Cross-domain eval and OOD/prediction probe suites used by the alignment matrix. Prompts and answer vocabularies only — the certified oracle and undecidable verdicts are stripped; score live against AmberTrace.
AT = gold — this dataset ships no answers
The AmberTrace verifier is the answer. Every certified-answer column
(gold / oracle / decision / triage_reason / undecidable) has been
stripped from these files: a public (features → certified decision) map
would give the decision function away and let any leaderboard be gamed.
Score live by submitting the prompts/features to AmberTrace and reading the
certificate at eval time — the verifier is the oracle, not a static key.
Files
| File | Rows | Fields | Notes |
|---|---|---|---|
decision_eval_v1.jsonl |
1350 | id, domain, prompt, vocabulary, difficulty |
1,350-item cross-domain decision eval; oracle + undecidable stripped. |
ood_probe_v1.jsonl |
120 | id, domain, prompt, vocabulary, difficulty |
120-item out-of-distribution probe; oracle + undecidable stripped. |
prediction_eval_v1.jsonl |
40 | id, domain, prompt, vocabulary, difficulty |
40-item symbolic-forecast eval; oracle + undecidable stripped. |
Provenance
These platforms were agent-authored via the AmberTrace SDK and
acceptance-gated; the prompt/feature files are generated reproducibly from
the source repo. See ambertrace-rlvr and its data/ +
examples/gen_*.py generators.
Reproduce / score
- Library:
pip install ambertrace-rlvr— https://pypi.org/project/ambertrace-rlvr/ - Source, methodology, and the live-scoring path: https://github.com/ambertrace-labs/ambertrace-rlvr
- Publishing policy (AT = gold):
docs/HUGGINGFACE.md
License
MIT (see the source repo).