{ "base_model": "Qwen/Qwen3-8B", "base_revision": "b968826d9c46dd6066d109eabc6255188de91218", "checkpoint_sha256": "c11034fea2a42e6b3a0e8ca1252cafa5e9e7419b2b750bd8308952e22bc36fbc", "method": "kveraser", "seed": 42, "dtype": "bf16", "attn_backend": "sdpa", "placement": "single_gpu", "warmup_samples": 3, "online_timing": "Erasing, query replay, and generation; excludes loading and initial prefill.", "environment": { "cuda": "12.4", "devices": { "0": { "name": "NVIDIA RTX 6000 Ada Generation", "total_bytes": 50951487488 } }, "peft": "0.18.0", "python": "3.10.20", "torch": "2.6.0+cu124", "transformers": "4.57.6" }, "results": { "eval/niah/ctx32000": { "em": 0.99, "f1": null, "gpu_memory": { "0": { "additional_peak_bytes": 546727936, "allocated_before_bytes": 21184952320, "online_peak_allocated_bytes": 21731680256, "sample_peak_allocated_bytes": 24601723904 } }, "mean_generated_tokens": 8.03, "online_time_s": { "mean": 0.5142248917417601, "p50": 0.5102429335238412, "p95": 0.5300618196662981, "total": 51.42248917417601 }, "samples": 100, "tokens_to_generate": 32 }, "eval/2wiki/default": { "em": 0.9069767441860465, "f1": 0.9163482834994462, "gpu_memory": { "0": { "additional_peak_bytes": 268969472, "allocated_before_bytes": 18794143232, "online_peak_allocated_bytes": 19063112704, "sample_peak_allocated_bytes": 20474419712 } }, "mean_generated_tokens": 7.531976744186046, "online_time_s": { "mean": 0.3682723147667698, "p50": 0.4042185049620457, "p95": 0.49807005331385873, "total": 126.68567627976881 }, "samples": 344, "tokens_to_generate": 32 }, "eval/musique/default": { "em": 0.8466666666666667, "f1": 0.8667089947089949, "gpu_memory": { "0": { "additional_peak_bytes": 272650752, "allocated_before_bytes": 18801221120, "online_peak_allocated_bytes": 19072865280, "sample_peak_allocated_bytes": 20484922880 } }, "mean_generated_tokens": 5.156666666666666, "online_time_s": { "mean": 0.2802521892004491, "p50": 0.26830165553838015, "p95": 0.3870559116388904, "total": 84.07565676013473 }, "samples": 300, "tokens_to_generate": 32 }, "eval/niah/ctx16000": { "em": 1.0, "f1": null, "gpu_memory": { "0": { "additional_peak_bytes": 281001984, "allocated_before_bytes": 18824815616, "online_peak_allocated_bytes": 19105817600, "sample_peak_allocated_bytes": 20527882752 } }, "mean_generated_tokens": 8.0, "online_time_s": { "mean": 0.4066749042732408, "p50": 0.40496365050785244, "p95": 0.4213551172724692, "total": 40.66749042732408 }, "samples": 100, "tokens_to_generate": 32 }, "eval/bfcl/default": { "em": 0.85, "f1": null, "gpu_memory": { "0": { "additional_peak_bytes": 515848704, "allocated_before_bytes": 21069217792, "online_peak_allocated_bytes": 21585066496, "sample_peak_allocated_bytes": 24401938432 } }, "mean_generated_tokens": 3.7, "online_time_s": { "mean": 0.2751596597000025, "p50": 0.26906770549248904, "p95": 0.35671693200129084, "total": 27.515965970000252 }, "samples": 100, "tokens_to_generate": 32 }, "eval/iirc/default": { "em": 0.9157303370786517, "f1": 0.9245585874799357, "gpu_memory": { "0": { "additional_peak_bytes": 270984192, "allocated_before_bytes": 18805350400, "online_peak_allocated_bytes": 19076334592, "sample_peak_allocated_bytes": 20494448640 } }, "mean_generated_tokens": 4.617977528089888, "online_time_s": { "mean": 0.2729848635430694, "p50": 0.24699980198056437, "p95": 0.46759864510095217, "total": 48.59130571066635 }, "samples": 178, "tokens_to_generate": 32 }, "eval/niah/ctx8000": { "em": 1.0, "f1": null, "gpu_memory": { "0": { "additional_peak_bytes": 147274240, "allocated_before_bytes": 17649986048, "online_peak_allocated_bytes": 17793384960, "sample_peak_allocated_bytes": 18499313152 } }, "mean_generated_tokens": 8.0, "online_time_s": { "mean": 0.3958572719770018, "p50": 0.39497024498996325, "p95": 0.40763832080410795, "total": 39.58572719770018 }, "samples": 100, "tokens_to_generate": 32 }, "eval/niah/ctx4000": { "em": 1.0, "f1": null, "gpu_memory": { "0": { "additional_peak_bytes": 78681600, "allocated_before_bytes": 17059311104, "online_peak_allocated_bytes": 17135731200, "sample_peak_allocated_bytes": 17479739904 } }, "mean_generated_tokens": 8.0, "online_time_s": { "mean": 0.3570333040604601, "p50": 0.35512368800118566, "p95": 0.3695398859825218, "total": 35.70333040604601 }, "samples": 100, "tokens_to_generate": 32 }, "eval/niah/ctx2000": { "em": 1.0, "f1": null, "gpu_memory": { "0": { "additional_peak_bytes": 47998464, "allocated_before_bytes": 16761558528, "online_peak_allocated_bytes": 16809556992, "sample_peak_allocated_bytes": 16965440000 } }, "mean_generated_tokens": 8.0, "online_time_s": { "mean": 0.35657244847039693, "p50": 0.3525064610003028, "p95": 0.3720976054901257, "total": 35.65724484703969 }, "samples": 100, "tokens_to_generate": 32 }, "eval/niah/ctx1000": { "em": 1.0, "f1": null, "gpu_memory": { "0": { "additional_peak_bytes": 29080064, "allocated_before_bytes": 16616574976, "online_peak_allocated_bytes": 16643975168, "sample_peak_allocated_bytes": 16717585408 } }, "mean_generated_tokens": 8.0, "online_time_s": { "mean": 0.3563199064537184, "p50": 0.3537737209990155, "p95": 0.3704430579848122, "total": 35.63199064537184 }, "samples": 100, "tokens_to_generate": 32 } } }