license: mit
language:

• en
• zh
  task_categories:
• text-classification
  tags:
• ngila
• evaluation
• refusal
  configs:
• config_name: default
  data_files:
  • split: test
    path: test.jsonl

ngila-ai/ngila-benchmarks-mit-v0.2

Version v0.2: normalized research records for independent measurement of refusal,
over-refusal, factuality and calibration. This release does not provide a single
ngila score or verified training targets.

Counts are records, not unique prompts: {"train": 0, "validation": 0, "test": 41231}. Sources: openaimod, harmbench, gsm8k, safetybench, simpleqa_verified, mmlu.
Some evaluation suites use upstream splits named train or validation; they are
reserved as test here. Original split names remain in source_split.

See NOTICE.md for attribution and sources.json for pinned revisions and input hashes.
Source labels apply only at label_scope; do not transfer response/QA labels to prompts.
Factual/math/MC answer keys, where available, are evaluation references only.
SafetyBench test answer keys are withheld; local test accuracy cannot be computed.
source_labels_json and reference_response_json are JSON-encoded to preserve
heterogeneous upstream schemas. prompt_risk=unknown remains unknown.
sft_target is null; these are prompt/annotation research data, not approved SFT answers.

Official test records remain test-only. Training/validation splits are deterministic
prompt/family groups. Exact cross-source overlap is excluded from training;
semantic paraphrase leakage has not been audited. Do not train on the test split.

This dataset contains sensitive benchmark material. Published behavior labels are
not assertions of factual truth. Assess generation outputs independently.