feat: add PR E benchmarks to STANDARD_TASK_MAP and model_roles support - #142
Merged
Conversation
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars. ⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Team Run ID: Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
ruivieira
reviewed
Sep 4, 2026
ruivieira
left a comment
ruivieira
left a comment
Member
There was a problem hiding this comment.
Thanks for the PR @j-mok-dev. LGTM, just left a minor comment.
| # models (e.g. HLE defaults to an OpenRouter judge). Provider YAML can | ||
| # specify parameters.model_roles: {grader: "openai/gpt-4o-mini"} to | ||
| # override at runtime without changing inspect-evals upstream defaults. | ||
| for role, spec in (config.parameters.get("model_roles") or {}).items(): |
Member
There was a problem hiding this comment.
Should we add an explicit isinstance guard for model_roles?
if a list is passed (eg model_roles: [openai/gpt-4o-mini]) wouldn't the truthy list pass the check and .items() raise AttributeError?
We could do
model_roles = config.parameters.get("model_roles") or {}
if not isinstance(model_roles, dict):
raise ValueError(
f"parameters.model_roles must be a dict (got {type(model_roles).__name__}). "
"Example: model_roles: {grader: openai/gpt-4o-mini}"
)
for role, spec in model_roles.items():
cmd += ["--model-role", f"{role}={spec}"]
Add HLE, BFCL, CyberSecEval-2-PI, DocVQA to STANDARD_TASK_MAP. Support parameters.model_roles in standard mode for judge/grader override (e.g. HLE's OpenRouter default → openai/gpt-4o-mini). Add 3 unit tests for model_roles injection. Refs: eval-hub#936, eval-hub#941 Signed-off-by: Jooyeon Mok <jmok@redhat.com>
j-mok-dev
force-pushed
the
feat/curated-collections-pr-e
branch
from
September 4, 2026 14:24
1b147ea to
7e7c186
Compare
ruivieira
approved these changes
Sep 4, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What and why
Add 4 new inspect-evals benchmarks to
STANDARD_TASK_MAP(HLE, BFCL, CyberSecEval-2-PI, DocVQA) and supportparameters.model_rolesin standard mode for judge/grader model override.The
model_rolesfeature allows benchmarks like HLE — which hardcodes an OpenRouter judge upstream — to be redirected toopenai/gpt-4o-mini(or any other model) at runtime via provider/collection YAML, without patching inspect-evals.Companion PR: eval-hub/eval-hub#954
Issue: eval-hub/eval-hub#941
Assisted-by: Cursor (Claude)
Type
Testing
Tests added or updated
Tested manually
Unit tests: 63/63 pass (3 new
model_rolestests)Local smoke: HLE, BFCL, CyberSecEval-PI, DocVQA PASS (Ollama + gpt-4o-mini judge)
lighteval/RULER benchmarks: blocked locally (Ollama tokenizer/API limits)
Reviewer: staging validation with vLLM endpoint needed for full coverage