Table 10 reports the full zero-shot evaluation row for Pythia-410M.

openai/gpt-5.6-luna-pro

eb72c786d2

Pile validation perplexity remains unresolved: lm-evaluation-harness 0.4.2 exposes Pile component tasks with a train split rather than the paper's validation split, and the available public metadata did not establish a matching validation artifact and evaluator. No published Pile value was copied. · The fixed paper and its Table 10 values were inspected during protocol interpretation. Those exposed source values are context only and are not presented as fresh evidence. · The fresh comparison uses lm-eval 0.4.2 ordinary acc, selected from the paper's explicit acc header and the evaluator's separate acc/acc_norm definitions. Historical evaluator, dataset, or preprocessing differences may affect direct numerical comparison; retained acc_norm values are an alternative field and were excluded from the reported aggregate. · Unresolved metric interpretation: m-t10-p410-pile: The fixed paper defines this as perplexity on the Pile validation split. The real lm-evaluation-harness 0.4.2 catalog exposes only component Pile tasks whose YAML sets `test_split: train`; its `pile` group has no validation evaluator. A metadata-only check of the public EleutherAI/pile artifact found configs but no usable split metadata and did not identify a validation artifact. Without a verified matching validation split and the paper's evaluator identity, this measurement remains unresolved and no value is substituted. · Unresolved measurements: m-t10-p410-pile: The preserved parser evidence did not contain a trustworthy metric: JSON 指针没有指向有限数值 · The Pile validation column was not independently evaluated, so the complete Table 10 row is not covered by demonstrated fresh computation. · The decisive full evaluation command and complete raw harness output are absent from the supplied non-truncated audit; 138 commands are omitted and some retained outputs are truncated. · The retained evaluator implementation snapshot is available, but the exact executed repository file is flagged as omitted or having a workspace-digest mismatch. · The source header uses acc, while several paper values are numerically closer to retained acc_norm fields; the historical metric mapping requires direct source/evaluator evidence not supplied. · The paper-reported average cannot be securely compared with the recomputed ordinary-acc average until the source operand definitions are resolved. · Execution evidence provenance is unknown: No supplied command record demonstrates the complete seven-task evaluation and its connection to the retained measurements. The available item_208 record constructs the parser artifact from full-harness/results.json but does not establish how that input was generated; other audit content is omitted or truncated. · Execution evidence provenance is unknown: No supplied command record demonstrates the complete seven-task evaluation and its connection to the retained measurements. The available item_208 record constructs the parser artifact from full-harness/results.json but does not establish how that input was generated; other audit content is omitted or truncated. Some original claim measurements remain without evidence; available measurements are assessed individually.