Repository navigation
Add FrontierScience Research benchmark - #1553
Merged
Merged
Conversation
Signed-off-by: Jiacheng Xu <jiachengx@nvidia.com>
Signed-off-by: Jiacheng Xu <jiachengx@nvidia.com>
Signed-off-by: Jiacheng Xu <jiachengx@nvidia.com>
jiacheng-xu
changed the title
Signed-off-by: Jiacheng Xu <jiachengx@nvidia.com>
Signed-off-by: Jiacheng Xu <jiachengx@nvidia.com>
jiacheng-xu
marked this pull request as ready for review
June 9, 2026 20:49
cmunley1
reviewed
Jun 9, 2026
| description: FrontierScience science-olympiad answer grading via single-pass YES/NO LLM judge | ||
| value: Evaluate free-form chemistry/biology/physics olympiad answers with judge-graded equivalence | ||
| description: FrontierScience answer grading via single-pass LLM judge | ||
| value: Evaluate FrontierScience Olympiad short answers or Research rubric-scored answers |
Contributor
There was a problem hiding this comment.
should the resources server remain generic (ie "free-form chemistry/biology/physics olympiad answers with judge-graded equivalence") while the benchmark is specific? (ie " FrontierScience Olympiad short answers or Research rubric-scored answers")
Just a nit, probably fine.
Contributor
Author
There was a problem hiding this comment.
Yeah I also thought about this.
Is there any rubric based resource server I can use directly? I know https://github.com/NVIDIA-NeMo/Gym/tree/main/resources_servers/equivalence_llm_judge but I am not sure if we want to overload it with "rubrics" eval.
Contributor
There was a problem hiding this comment.
i am not sure, but would be nice to have one
Contributor
|
/ok to test 7d9aeba |
wprazuch
pushed a commit
that referenced
this pull request
Jun 22, 2026
## Summary Adds a `frontierscience_research` benchmark alongside the existing FrontierScience Olympiad benchmark. - Downloads `openai/frontierscience` `research/test.jsonl` from Hugging Face and prepares 60 research rows. - Configures the benchmark to use the existing FrontierScience judge resource server with `judge_mode: research`. - Extends `frontierscience_judge` with 10-point rubric scoring, normalized rubric metrics, thresholded rewards, and reusable prompt-path resolution. - Adds a Research judge prompt, docs, README table update, and focused tests for rubric parsing and scoring behavior. ## Validation - `.venv/bin/python -m pytest resources_servers/frontierscience_judge/tests/test_app.py -q` (`45 passed`) - `.venv/bin/ng_prepare_benchmark '+config_paths=[benchmarks/frontierscience_research/config.yaml]'` - `.venv/bin/ng_prepare_benchmark '+config_paths=[benchmarks/frontierscience_olympiad/config.yaml]'` - `.venv/bin/ng_list_benchmarks` - `git diff main...HEAD -- benchmarks/frontierscience_olympiad` (empty) - Reproduced the 60-task Research run via NVIDIA inference API: Claude Opus 4.5 high reached 16.67% accuracy (paper: 17.5%); OpenAI GPT-5.1 high reached 21.67% on all rows, with 59/60 completed at the primary setting (paper: 19.4%). External API smoke/run artifacts were kept under ignored `results/` paths and are not committed. --------- Signed-off-by: Jiacheng Xu <jiachengx@nvidia.com> Co-authored-by: Christian Munley <cmunley@nvidia.com>
This branch was previously deployed
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
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.
Summary
Adds a
frontierscience_researchbenchmark alongside the existing FrontierScience Olympiad benchmark.openai/frontierscienceresearch/test.jsonlfrom Hugging Face and prepares 60 research rows.judge_mode: research.frontierscience_judgewith 10-point rubric scoring, normalized rubric metrics, thresholded rewards, and reusable prompt-path resolution.Validation
.venv/bin/python -m pytest resources_servers/frontierscience_judge/tests/test_app.py -q(45 passed).venv/bin/ng_prepare_benchmark '+config_paths=[benchmarks/frontierscience_research/config.yaml]'.venv/bin/ng_prepare_benchmark '+config_paths=[benchmarks/frontierscience_olympiad/config.yaml]'.venv/bin/ng_list_benchmarksgit diff main...HEAD -- benchmarks/frontierscience_olympiad(empty)External API smoke/run artifacts were kept under ignored
results/paths and are not committed.