[RikiConrey]

Summary

Add Safety 1.2.1 scoring with these updates from Safety 1.2, bringing scoring in line with Security 1.1:

• Use continuous evaluator violation probabilities instead of binary safety labels.
• Use refitted prompt difficulty and discrimination parameters (a/b) to match the calibration in Security 1.1 based on the new 19 SUT list.
• Integrate the revised probability evaluator; its improved centering and spread and limited ground-truth calibration are supplied by the evaluator service.
• Export the new score payload: overall and per-hazard reference percentile scores, mean violation probabilities, null grades, scored-response counts, and the parameter-file digest.

Retain Safety 1.2's requirement that every expected response have a successful evaluation. Any missing or invalid evaluation blocks publication.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[dhosterman]

This fails with:

❯ uv run modelbench benchmark general -v 1.2.1 -s demo_yes_no -m 1
      Built modelbench @ file:///Users/daniel/dev/src/github.com/mlcommons/modelbench                                                                                                                                 Uninstalled 1 package in 1ms
Installed 1 package in 1ms
Traceback (most recent call last):
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/src/modelgauge/instance_factory.py", line 85, in _get_entry
    entry = self._lookup[uid]
            ~~~~~~~~~~~~^^^^^
KeyError: 'safety-v1.3'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/.venv/bin/modelbench", line 10, in <module>
    sys.exit(cli())
             ~~~^^
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/.venv/lib/python3.14/site-packages/click/core.py", line 1569, in __call__
    return self.main(*args, **kwargs)
           ~~~~~~~~~^^^^^^^^^^^^^^^^^
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/.venv/lib/python3.14/site-packages/click/core.py", line 1490, in main
    rv = self.invoke(ctx)
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/.venv/lib/python3.14/site-packages/click/core.py", line 1970, in invoke
    return _process_result(sub_ctx.command.invoke(sub_ctx))
                           ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/.venv/lib/python3.14/site-packages/click/core.py", line 1970, in invoke
    return _process_result(sub_ctx.command.invoke(sub_ctx))
                           ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/.venv/lib/python3.14/site-packages/click/core.py", line 1353, in invoke
    return ctx.invoke(self.callback, **ctx.params)
           ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/.venv/lib/python3.14/site-packages/click/core.py", line 907, in invoke
    return callback(*args, **kwargs)
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/src/modelbench/cli.py", line 135, in wrapper
    return func(*args, **kwargs)
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/.venv/lib/python3.14/site-packages/click/decorators.py", line 34, in new_func
    return f(get_current_context(), *args, **kwargs)
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/src/modelbench/cli.py", line 244, in general_benchmark
    check_benchmark(benchmark)
    ~~~~~~~~~~~~~~~^^^^^^^^^^^
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/src/modelbench/cli.py", line 389, in check_benchmark
    check_secrets(secrets, test_uids=test_uids)
    ~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/src/modelgauge/preflight.py", line 35, in check_secrets
    missing_secrets.extend(get_missing_secrets(secrets, ANNOTATORS, test_cls.get_annotators()))
                           ~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/src/modelgauge/preflight.py", line 21, in get_missing_secrets
    missing_secrets.extend(registry.get_missing_dependencies(uid, secrets=secrets))
                           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/src/modelgauge/instance_factory.py", line 78, in get_missing_dependencies
    entry = self._get_entry(uid)
  File "/Users/daniel/dev/src/github.com/mlcommons/modelbench/src/modelgauge/instance_factory.py", line 87, in _get_entry
    raise ValueError(f"No registration for {uid}. Known uids:\n{self.compact_uid_list()}")
ValueError: No registration for safety-v1.3. Known uids:
        demo_annotator
        demo_goody_annotator
        llama_guard_1
        llama_guard_2
        safety-v1.1
        safety-v1.2
        security-v1.0.2
        security-v1.1

Looks like maybe a registration issue?

[RikiConrey]

@dhosterman You're right—the registration for safety-v1.3 was missing. Fixed and pushed in 2460dd8. Added a regression test covering benchmark preflight and evaluator lookup; all 95 Safety/Security scorer and Cheval annotator tests pass.