{ "schemaVersion": 1, "createdUtc": "2026-08-26T16:55:41.044394+00:00", "experiment": "UD-Q2_K_XL-Unsloth", "status": "complete", "models": { "native": { "path": "/Qwen3.8-27B-Quark-AWQ-MXFP4-native.gguf", "bytes": 18892854880, "decimalGB": 18.89285488, "GiB": 17.595342248678207, "sha256": "1a37c48570811215fa1a9d2a493211293ca375d17287771e85d28642946e88fc" }, "candidate": { "path": "/Qwen3.8-27B-Quark-MXFP4-UD-Q2_K_XL-Unsloth.gguf", "bytes": 9809893280, "decimalGB": 9.80989328, "GiB": 9.136175066232681, "sha256": "9b899d16649813748689f2b99407b299cc9eeeae66289cf06066b050f10f3675" }, "savings": { "bytes": 9082961600, "decimalGB": 9.0829616, "GiB": 8.459167182445526, "percentOfNative": 48.07617301721422 } }, "tensorAudit": { "verificationPassed": true, "tensorCount": 866, "changedTensorCount": 369, "unchangedPayloadsVerifiedByteExact": 497, "nativeMxfp4PreservedCount": 137, "nativeMxfp4TensorCount": 496, "effectiveTypeCounts": { "F32": 360, "IQ1_M": 1, "IQ1_S": 20, "IQ2_S": 67, "IQ2_XS": 34, "IQ2_XXS": 48, "IQ3_S": 57, "IQ3_XXS": 112, "MXFP4": 137, "Q2_K": 16, "Q3_K": 5, "Q4_K": 1, "Q6_K": 6, "Q8_0": 2 } }, "benchmark": { "domain": "general", "tokenTarget": 32768, "chunks": 15, "context": 2048, "threads": 4, "gpuLayers": 0, "gpuVisible": false, "nativeStandalonePpl": 5.8033, "nativeStandalonePplError": 0.11146, "paired": { "candidatePpl": 6.280284, "basePpl": 5.801511, "pplDifference": 0.478773, "pplRatio": 1.082526, "logPplRatio": 0.079297, "correlationPercent": 97.36 }, "kld": { "mean": 0.111015, "standardError": 0.001835, "maximum": 5.839242, "p99_9": 2.934114, "p99": 0.993161, "p95": 0.397965, "p90": 0.247337, "median": 0.052314 }, "tokenProbability": { "meanDeltaPercent": -2.07, "rmsDeltaPercent": 9.66, "sameTopPercent": 86.087 }, "scope": { "languageLogits": true, "visionProjectorEvaluated": false, "mtpTensorsEvaluated": false, "note": "llama-perplexity logged block 64 MTP tensors as unused; ordinary KLD primarily measures the changed token embedding and output tensors" } }, "imatrixApplicability": { "supplied": true, "label": "Unsloth", "sha256": "0ee5b10bd0c2fa2127c6f4b43dbfe1efd71e383b63217af9dade1de36599f1c1", "entries": 496, "changedTensorsWithEntries": 359, "changedTensorCount": 369, "note": "Coverage is inferred from llama.cpp's per-conversion missing-weight messages; tensors present in the imatrix are supplied to the selected quantizer." }, "artifacts": { "corpus": { "path": "/Experiments/MQ-IQ4_XS_1-Generic/benchmark/_ppl_corpora/ppl_corpus_general.txt", "bytes": 131514, "sha256": "5d38d98dce15f54e9a1a926187b6058e65cd8b3dd9b5cc2729b0a5fd249228b0" }, "referenceLogits": { "path": "/Experiments/MQ-IQ4_XS_1-Generic/benchmark/native-reference/logits/kld_logits_general.bin", "bytes": 7621186460, "sha256": "604db2df7253cba18a6c0a569ed819ee3065d4ac20cd94ae72f4329e7bba8516" }, "nativeLog": { "path": "/Experiments/MQ-IQ4_XS_1-Generic/benchmark/native-reference/perplexity_general.log", "sha256": "bade3445565a354969f7515045507b854bee83f6695eec7490b70eab9c6eebc2" }, "candidateLog": { "path": "/Experiments/UD-Q2_K_XL-Unsloth/benchmark/candidate/perplexity_general.log", "sha256": "245d6d9f1551cc03b74f98e7c20fb8372098c7e077fd71d0f517f48cc7b47a35" }, "quantizationLog": { "path": "/Experiments/UD-Q2_K_XL-Unsloth/quantization.log", "sha256": "99260ff506669dcc6eae299f00260c2c4ceebbc61c6c204a72a0dc7b696c35d8" } } }