Table 10 reports the full zero-shot evaluation row for RWKV4-1.5B.

openai/gpt-5.6-luna-pro

1f72b64a0a

Pile validation perplexity was not reported as source-equivalent because its aggregation policy is unresolved in the fixed paper sources and no matching lm-evaluation-harness task was available. · The numerical evaluation was run on the allocated NVIDIA L4; this result does not establish any paper-specific hardware or speed claim. · The result covers the assigned RWKV4-1.5B row, not the wider Table 10 model catalog. · Unresolved metric interpretation: m-t10-rwkv15-pile: The paper resolves the quantity as perplexity on the Pile validation split with the GPT-NeoX tokenizer, but the cited source does not specify the document-boundary/sequence-packing aggregation used to obtain the validation negative log likelihood, and lm-evaluation-harness 0.4.2 has no Pile-validation task. A prepared evaluator therefore records the policy explicitly and the GPU phase must resolve or disclose this remaining condition before treating the value as source-equivalent. · Unresolved measurements: m-t10-rwkv15-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 executed, so the full RWKV4-1.5B source row remains uncovered. · The declared workload spans the broader Table 10 catalog, but captured execution covers only the assigned RWKV4-1.5B row. · The cited README and evaluator source needed to resolve ordinary versus normalized accuracy were not retained in evidenceContext. · The pinned checkpoint hash establishes the evaluated artifact, but not exact equivalence to the paper-side checkpoint or training-stage state. · The L4 run establishes these evaluation results on that device; it does not establish any paper-specific hardware or speed claim. · The captured CUDA run demonstrates a fresh seven-task evaluation of the pinned RWKV4-1.5B artifact and a traceable recomputed average. LAMBADA perplexity and accuracy, and WinoGrande accuracy, are approximately reproduced. However, HellaSwag, PIQA, ARC-E, ARC-C, and OpenBookQA each expose both ordinary accuracy and normalized accuracy in the raw evaluator output, while the retained paper excerpt labels the columns only as generic accuracy. The cited README/evaluator context needed to prove the paper-to-field mapping was not retained. Therefore neither numeric proximity nor discrepancy can establish equivalence or contradiction for those quantities. The Pile validation column was not executed, so the full source row was not reproduced. Some original claim measurements remain without evidence; available measurements are assessed individually.