When Strong Reasoners Join the Pool: run record and code (NVIDIA B300, MATH-500)
收藏资源简介:
Run record and code for the technical report "When Strong Reasoners Join the Pool: Small Union Premiums and Costly Runaway Generations in LLM Routing". All 500 MATH-500 problems were answered 30 times each by Llama-3.3-70B-Instruct, NVIDIA Nemotron Nano 9B v2 and Nemotron Super 49B v1.5 under a protocol fixed before generation (vendor sampling, 65,536-token cap), and the 45,000 responses were scored with a pinned Math-Verify 0.9.0 scorer. The archive contains the pre-commitment and its amendment, seed manifests, an append-only hash-chained attempt ledger, engine and hardware receipts, every response with its score, a quarantined Llama-3.1-8B attempt, the platform record, the generation/scoring/analysis code with tests, and the report source. One script regenerates every number, table and figure of the report from the archived record. Host IP addresses and hardware serial numbers were removed from operational logs; all hash-bound files are byte-identical to the originals. Code: MIT. Run record and report: CC BY 4.0. Model responses remain subject to the Llama 3.3 Community License and the NVIDIA Open Model License (see NOTICE.md).



