{ "experiment": "D82", "prereg_sha256": "856f0f237dbedb6ba51a52ab028b5ba0f6044edab14752ca7dbec01f441398e9", "sets": { "SEL-82": { "set_sha256": "7fb4c3d16d3f1a4a72e52b2e25fd9385d9b4ff04165aabcdd0d350e2a2cf8b3c", "verdict": "H0_SELECTED", "validity": "VALID", "n_paired_items": 10314, "n_gated_items": 10198, "axes_n": { "ko": 3695, "en": 1935, "number": 3328, "reasoning": 1240, "ko_indomain": 116 }, "underpowered_prereg": [ "UNDERPOWERED_number", "UNDERPOWERED_reasoning" ], "note": "items are not distributed; the sha256 identifies the frozen set" }, "FINAL-82": { "set_sha256": "8088e07dc3746781eacacbe715c42666462f92f7caf1507345cbff27c40efca1", "verdict": "FINAL82_H0_PASS", "validity": "VALID", "n_paired_items": 10496, "n_gated_items": 10210, "axes_n": { "ko": 3696, "en": 1942, "number": 3322, "reasoning": 1250, "ko_math": 169, "ko_indomain": 117 }, "underpowered_prereg": [ "UNDERPOWERED_number", "UNDERPOWERED_reasoning" ], "note": "items are not distributed; the sha256 identifies the frozen set" } }, "non_inferiority": { "comparison": "H0 − A81, gate_all(ko·en·number·reasoning 게이트 축 문항 전부) 문항 단위 paired 정답 차이", "margin_pp": -1.5, "alpha_one_sided": 0.025, "z": 1.959964, "k_confirmatory": 1 }, "superiority_test": "none preregistered; no superiority claim", "gates_vs_bf16": { "dacc": -0.03, "format_pp": 0.02, "ko_ret": 0.9, "mem_gib_below": 5.98, "num_dacc": -0.03, "reasoning_dacc": -0.03, "tier": 0.9, "tpot_vs_bf16": 1.1 }, "sources": [ "KMMLU (evaluation only)", "KoBEST", "MMLU", "HellaSwag", "BBH", "BIG-bench", "HRM8K", "programmatic numeric generator (new seeds)" ], "excluded": "every source item used in DEV-80, SEL-81, FINAL-81 and REL-81; SEL-82 and FINAL-82 share no items or source texts", "aihub_data": "none", "environment": { "gpu": "1x NVIDIA B200", "vllm": "0.29.0", "torch": "2.13.0+cu130", "VLLM_BATCH_INVARIANT": "unset", "decoding": "greedy, enable_thinking=False, max_model_len 8192, max_num_seqs 8, seed 0" }, "serving_batch1": "32 prompts, 256 output tokens, same GPU", "concurrency": { "method": "offline vLLM LLM.generate; first 32 FINAL-82 numeric generation items; 256 fixed output tokens (ignore_eos); greedy; 1 warmup + median of 3 runs; one pod, arms run sequentially; no HTTP server", "role": "report-only, not used for selection" }, "gpu_hours_d82_total": 3.186, "related": { "model": "ThakiCloud/Qwen3-8B-NVFP4A16-INT8Head-B200", "control_model": "ThakiCloud/Qwen3-8B-NVFP4A16-INT8-Mixed-B200", "d81_results": "ThakiCloud/B200-LowBit-Quantization-Benchmarks" } }