Add Google FACTS Grounding benchmark#1407
Draft
MahanFathi wants to merge 9 commits into
Draft
Conversation
Signed-off-by: Mahan Fathi <mfathi@nvidia.com>
Mirrors google-facts reference compute_scores: per-judge grounding/quality verdicts fanned out across Gemini 3.1 Pro / GPT-5.2 / Claude Opus 4.5 (all via inference-api.nvidia.com), aggregated into unadjusted_factuality, final_factuality (consensus-ineligible zeroed), eligibility_rate, per-judge slices, Wilson 95% CIs, and sentence-level label micro-averages. Handles per-endpoint quirks: drops temperature for GPT-5/o-series, drops top_p for non-Gemini judges (Bedrock rejects both). Single-judge fallback preserved when judge_models is empty. Signed-off-by: Mahan Fathi <mfathi@nvidia.com>
When judges emit a single JSON array ``[{...}, {...}]`` instead of the
newline-delimited object form, ``parse_grounding_json`` crashed with
``AttributeError: 'list' object has no attribute 'get'``. Flatten either
shape into a list of dicts; skip non-dict entries defensively.
This happened to surface on the Nemotron v2 run (parse_reasoning=True →
shorter, cleaner final answers → judges more likely to return a compact
single-array response), crashing the whole judge stage after 45s.
Signed-off-by: Mahan Fathi <mfathi@nvidia.com>
Signed-off-by: Mahan Fathi <mfathi@nvidia.com>
Signed-off-by: Mahan Fathi <mfathi@nvidia.com>
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 support for the Google FACTS Grounding benchmark in NeMo-Skills.
This includes:
google/FACTS-grounding-publicfinal_factuality,unadjusted_factuality, eligibility rate, confidence intervals, per-judge scores, and sentence-label statisticsValidation
pre-commit run --files nemo_skills/dataset/facts_grounding/__init__.py nemo_skills/dataset/facts_grounding/prepare.py nemo_skills/evaluation/metrics/facts_grounding_metrics.py nemo_skills/evaluation/metrics/map_metrics.py nemo_skills/inference/eval/facts_grounding_judge.py nemo_skills/prompt/config/generic/facts_grounding.yaml nemo_skills/prompt/config/judge/facts_grounding.yaml tests/test_facts_grounding_metrics.pypre-commit run --files docs/evaluation/other-benchmarks.mdpytest tests/test_facts_grounding_metrics.py -qpython -m py_compile nemo_skills/inference/eval/facts_grounding_judge.pyNotes
The local public-split run for Nemotron-3-Nano produced
final_factuality = 39.81%using the NeMo-Skills default judge set. This is a public-split comparison and is not directly identical to the Kaggle private leaderboard score.