10 random (doc, eval) pairs sampled (seed=42) from eval_to_docs.parquet of the MMLU vs nemotron_cc_math_v1 run (writeup).
- corpus:
gs://marin-us-east5/normalized/nemotron_cc_math_v1/4plus_b05688a8/outputs/main/ - evals:
gs://marin-us-east1/decontamination/mmlu-9fbdd5/cais/ - analysis:
gs://marin-us-east5/tmp/ttl=7d/rav/decon-v0/mmlu-vs-nemotron-math-v1/analysis/eval_to_docs.parquet
Each example shows the matched MMLU question (with answer) and the raw Nemotron document text it leaked into.