{ "arms": { "q-q3_K_M": { "card_memory_total_mib": "24576", "ended": "2026-10-15T11:01:53Z", "failed": null, "idle_preload": { "board": { "cards": [ "" ], "joules": 395.0088771033287, "mean_watts": 28.170651476808573, "peak_watts": 28.5, "per_card": { "": { "joules": 395.0088771033287, "mean_watts": 28.170651476808573, "min_watts": 27.99, "peak_watts": 28.5, "samples": 15, "seconds": 14.022000074386597 } }, "samples": 15, "seconds": 14.022000074386597 }, "wall": { "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "joules": 1371.6340050697327, "mean_watts": 105.4616331419642, "median_s_between_changes": null, "min_watts": 99.0, "peak_watts": 107.0, "samples": 14, "seconds": 13.00600004196167, "value_changes": 1 }, "window_epoch": [ 1792061927.12, 1792061941.142 ] }, "model": "mistral-nemo:12b-instruct-2407-q3_K_M", "num_ctx": 4096, "ollama_version": "0.32.13", "pin": { "fit": { "growth_bytes": 6902775808.0, "ps_size": 6579583712, "ps_size_vram": 6579583712, "ps_whole": true, "rule": "WHOLE when the named board grew by at least 90 %% of the store's bytes for this tag (weights + projector/draft; KV rides on top) AND /api/ps does not report a CPU share; ps alone is corroboration, never the verdict", "store_bytes": 6083102669, "whole": true }, "growth_mib": { "": 0.0, "": 6583.0 }, "ok": true }, "rc": 0, "refused": null, "started": "2026-10-15T10:58:41Z", "tasks": [ "c", "d" ], "think": "off", "think_budget": 8, "ups_stop": null }, "q-q4_K_M": { "card_memory_total_mib": "24576", "ended": "2026-10-15T05:44:55Z", "failed": null, "idle_preload": { "board": { "cards": [ "" ], "joules": 392.94713947534564, "mean_watts": 28.02761341012339, "peak_watts": 28.28, "per_card": { "": { "joules": 392.94713947534564, "mean_watts": 28.02761341012339, "min_watts": 27.75, "peak_watts": 28.28, "samples": 15, "seconds": 14.019999980926514 } }, "samples": 15, "seconds": 14.019999980926514 }, "wall": { "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "joules": 1352.6240043640137, "mean_watts": 104.0, "median_s_between_changes": null, "min_watts": 104.0, "peak_watts": 104.0, "samples": 14, "seconds": 13.00600004196167, "value_changes": 0 }, "window_epoch": [ 1792042910.26, 1792042924.28 ] }, "model": "mistral-nemo:12b-instruct-2407-q4_K_M", "num_ctx": 4096, "ollama_version": "0.32.13", "pin": { "fit": { "growth_bytes": 8207204352.0, "ps_size": 7884568001, "ps_size_vram": 7884568001, "ps_whole": true, "rule": "WHOLE when the named board grew by at least 90 %% of the store's bytes for this tag (weights + projector/draft; KV rides on top) AND /api/ps does not report a CPU share; ps alone is corroboration, never the verdict", "store_bytes": 7477217229, "whole": true }, "growth_mib": { "": 0.0, "": 7827.0 }, "ok": true }, "rc": 0, "refused": null, "started": "2026-10-15T05:41:44Z", "tasks": [ "c", "d" ], "think": "off", "think_budget": 8, "ups_stop": null }, "q-q5_K_M": { "card_memory_total_mib": "24576", "ended": "2026-10-15T10:49:32Z", "failed": null, "idle_preload": { "board": { "cards": [ "" ], "joules": 390.77918148756027, "mean_watts": 27.874968257284852, "peak_watts": 28.03, "per_card": { "": { "joules": 390.77918148756027, "mean_watts": 27.874968257284852, "min_watts": 27.75, "peak_watts": 28.03, "samples": 15, "seconds": 14.019000053405762 } }, "samples": 15, "seconds": 14.019000053405762 }, "wall": { "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "joules": 1313.4029948711395, "mean_watts": 100.99992310615774, "median_s_between_changes": 10.003000020980835, "min_watts": 100.0, "peak_watts": 102.0, "samples": 14, "seconds": 13.003999948501587, "value_changes": 2 }, "window_epoch": [ 1792061185.806, 1792061199.825 ] }, "model": "mistral-nemo:12b-instruct-2407-q5_K_M", "num_ctx": 4096, "ollama_version": "0.32.13", "pin": { "fit": { "growth_bytes": 9373220864.0, "ps_size": 9051108801, "ps_size_vram": 9051108801, "ps_whole": true, "rule": "WHOLE when the named board grew by at least 90 %% of the store's bytes for this tag (weights + projector/draft; KV rides on top) AND /api/ps does not report a CPU share; ps alone is corroboration, never the verdict", "store_bytes": 8727644109, "whole": true }, "growth_mib": { "": 0.0, "": 8939.0 }, "ok": true }, "rc": 0, "refused": null, "started": "2026-10-15T10:46:19Z", "tasks": [ "c", "d" ], "think": "off", "think_budget": 8, "ups_stop": null }, "q-q6_K": { "card_memory_total_mib": "24576", "ended": "2026-10-15T10:36:35Z", "failed": null, "idle_preload": { "board": { "cards": [ "" ], "joules": 361.5910344898701, "mean_watts": 25.791086660613125, "peak_watts": 26.01, "per_card": { "": { "joules": 361.5910344898701, "mean_watts": 25.791086660613125, "min_watts": 25.66, "peak_watts": 26.01, "samples": 15, "seconds": 14.019999980926514 } }, "samples": 15, "seconds": 14.019999980926514 }, "wall": { "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "joules": 1531.7080035209656, "mean_watts": 117.76933711972686, "median_s_between_changes": null, "min_watts": 97.0, "peak_watts": 133.0, "samples": 14, "seconds": 13.00600004196167, "value_changes": 1 }, "window_epoch": [ 1792060406.244, 1792060420.264 ] }, "model": "mistral-nemo:12b-instruct-2407-q6_K", "num_ctx": 4096, "ollama_version": "0.32.13", "pin": { "fit": { "growth_bytes": 10612637696.0, "ps_size": 10290557091, "ps_size_vram": 10290557091, "ps_whole": true, "rule": "WHOLE when the named board grew by at least 90 %% of the store's bytes for this tag (weights + projector/draft; KV rides on top) AND /api/ps does not report a CPU share; ps alone is corroboration, never the verdict", "store_bytes": 10056222667, "whole": true }, "growth_mib": { "": 0.0, "": 10121.0 }, "ok": true }, "rc": 0, "refused": null, "started": "2026-10-15T10:33:20Z", "tasks": [ "c", "d" ], "think": "off", "think_budget": 8, "ups_stop": null }, "q-q8_0": { "card_memory_total_mib": "24576", "ended": "2026-10-15T05:59:14Z", "failed": null, "idle_preload": { "board": { "cards": [ "" ], "joules": 391.70502854824065, "mean_watts": 27.943003640364914, "peak_watts": 28.23, "per_card": { "": { "joules": 391.70502854824065, "mean_watts": 27.943003640364914, "min_watts": 27.69, "peak_watts": 28.23, "samples": 15, "seconds": 14.01800012588501 } }, "samples": 15, "seconds": 14.01800012588501 }, "wall": { "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "joules": 1324.0095117092133, "mean_watts": 101.8077278014797, "median_s_between_changes": null, "min_watts": 101.0, "peak_watts": 102.0, "samples": 14, "seconds": 13.005000114440918, "value_changes": 1 }, "window_epoch": [ 1792043765.972, 1792043779.99 ] }, "model": "mistral-nemo:12b-instruct-2407-q8_0", "num_ctx": 4096, "ollama_version": "0.32.13", "pin": { "fit": { "growth_bytes": 13416529920.0, "ps_size": 13094187171, "ps_size_vram": 13094187171, "ps_whole": true, "rule": "WHOLE when the named board grew by at least 90 %% of the store's bytes for this tag (weights + projector/draft; KV rides on top) AND /api/ps does not report a CPU share; ps alone is corroboration, never the verdict", "store_bytes": 13022382027, "whole": true }, "growth_mib": { "": 0.0, "": 12795.0 }, "ok": true }, "rc": 0, "refused": null, "started": "2026-10-15T05:55:59Z", "tasks": [ "c", "d" ], "think": "off", "think_budget": 8, "ups_stop": null } }, "generated": "2026-10-15T11:01:53Z", "parser": "v2-benchbox-1", "probes": [ { "arm": "q-q3_K_M", "decode": { "board_j_per_1k_gen_tokens": 3677.8, "board_j_per_gen_token_phase": 3.6778, "board_j_per_gen_token_window": 4.2687, "board_j_window": 3278.3, "board_mean_w": 273.0, "board_net_j_per_gen_token_phase": 1.9607, "board_peak_w": 300.43, "decode_tok_s_median": 74.18, "file": "q-q3_K_M/probes.json + q-q3_K_M/probe-decode.board.csv", "gen_tokens_per_board_j": 0.272, "runs": 3, "tok_s_per_board_w": 0.2719, "tok_s_per_wall_w": 0.2168, "tokens": 768, "wall_j_per_gen_token_phase": 4.6136, "wall_mean_w": 342.5, "window_s": 12.01 }, "idle_resident_w": 127.5, "model": "mistral-nemo:12b-instruct-2407-q3_K_M", "prefill": { "board_j_per_prompt_token_phase": 0.07624, "board_j_per_prompt_token_window": 0.14672, "board_mean_w": 227.4, "board_peak_w": 257.83, "file": "q-q3_K_M/probes.json + q-q3_K_M/probe-prefill.board.csv", "prefill_tok_s_median": 2982.9, "prompt_tokens": 20159, "prompt_tokens_per_board_j": 6.82, "prompt_tokens_per_run_median": 2240, "runs": 9, "wall_j_per_prompt_token_window": 0.21577, "wall_mean_w": 362.4, "window_s": 13.01 }, "think": "off", "ttft": { "how": "the first 12 task_c prompts, run.py's body with stream true, in task order (so the prompt-cache condition is the task's own); TTFT = client clock on localhost from the request to the first chunk carrying content or thinking", "median_s": 0.5552, "n": 12, "p95_s": 0.6118 } }, { "arm": "q-q4_K_M", "decode": { "board_j_per_1k_gen_tokens": 3051.2, "board_j_per_gen_token_phase": 3.0512, "board_j_per_gen_token_window": 3.813, "board_j_window": 2928.4, "board_mean_w": 266.0, "board_net_j_per_gen_token_phase": 1.5765, "board_peak_w": 301.27, "decode_tok_s_median": 87.22, "file": "q-q4_K_M/probes.json + q-q4_K_M/probe-decode.board.csv", "gen_tokens_per_board_j": 0.328, "runs": 3, "tok_s_per_board_w": 0.3277, "tok_s_per_wall_w": 0.2992, "tokens": 768, "wall_j_per_gen_token_phase": 3.3423, "wall_mean_w": 291.3, "window_s": 11.01 }, "idle_resident_w": 128.5, "model": "mistral-nemo:12b-instruct-2407-q4_K_M", "prefill": { "board_j_per_prompt_token_phase": 0.06968, "board_j_per_prompt_token_window": 0.14041, "board_mean_w": 224.5, "board_peak_w": 250.96, "file": "q-q4_K_M/probes.json + q-q4_K_M/probe-prefill.board.csv", "prefill_tok_s_median": 3222.6, "prompt_tokens": 22399, "prompt_tokens_per_board_j": 7.12, "prompt_tokens_per_run_median": 2240.0, "runs": 10, "wall_j_per_prompt_token_window": 0.22795, "wall_mean_w": 392.7, "window_s": 14.01 }, "think": "off", "ttft": { "how": "the first 12 task_c prompts, run.py's body with stream true, in task order (so the prompt-cache condition is the task's own); TTFT = client clock on localhost from the request to the first chunk carrying content or thinking", "median_s": 0.5629, "n": 12, "p95_s": 0.5983 } }, { "arm": "q-q5_K_M", "decode": { "board_j_per_1k_gen_tokens": 3411.1, "board_j_per_gen_token_phase": 3.4111, "board_j_per_gen_token_window": 4.1919, "board_j_window": 3219.4, "board_mean_w": 268.1, "board_net_j_per_gen_token_phase": 1.7899, "board_peak_w": 301.84, "decode_tok_s_median": 78.59, "file": "q-q5_K_M/probes.json + q-q5_K_M/probe-decode.board.csv", "gen_tokens_per_board_j": 0.293, "runs": 3, "tok_s_per_board_w": 0.2932, "tok_s_per_wall_w": 0.2151, "tokens": 768, "wall_j_per_gen_token_phase": 4.6499, "wall_mean_w": 365.5, "window_s": 12.01 }, "idle_resident_w": 127.4, "model": "mistral-nemo:12b-instruct-2407-q5_K_M", "prefill": { "board_j_per_prompt_token_phase": 0.0716, "board_j_per_prompt_token_window": 0.14444, "board_mean_w": 223.8, "board_peak_w": 253.5, "file": "q-q5_K_M/probes.json + q-q5_K_M/probe-prefill.board.csv", "prefill_tok_s_median": 3126.1, "prompt_tokens": 20159, "prompt_tokens_per_board_j": 6.92, "prompt_tokens_per_run_median": 2240, "runs": 9, "wall_j_per_prompt_token_window": 0.24626, "wall_mean_w": 413.6, "window_s": 13.01 }, "think": "off", "ttft": { "how": "the first 12 task_c prompts, run.py's body with stream true, in task order (so the prompt-cache condition is the task's own); TTFT = client clock on localhost from the request to the first chunk carrying content or thinking", "median_s": 0.5649, "n": 12, "p95_s": 0.5864 } }, { "arm": "q-q6_K", "decode": { "board_j_per_1k_gen_tokens": 4138.8, "board_j_per_gen_token_phase": 4.1388, "board_j_per_gen_token_window": 4.9701, "board_j_window": 3817.0, "board_mean_w": 272.5, "board_net_j_per_gen_token_phase": 2.3545, "board_peak_w": 305.16, "decode_tok_s_median": 65.8, "file": "q-q6_K/probes.json + q-q6_K/probe-decode.board.csv", "gen_tokens_per_board_j": 0.242, "runs": 3, "tok_s_per_board_w": 0.2416, "tok_s_per_wall_w": 0.1679, "tokens": 768, "wall_j_per_gen_token_phase": 5.9543, "wall_mean_w": 392.0, "window_s": 14.01 }, "idle_resident_w": 117.5, "model": "mistral-nemo:12b-instruct-2407-q6_K", "prefill": { "board_j_per_prompt_token_phase": 0.07801, "board_j_per_prompt_token_window": 0.14491, "board_mean_w": 224.5, "board_peak_w": 262.57, "file": "q-q6_K/probes.json + q-q6_K/probe-prefill.board.csv", "prefill_tok_s_median": 2876.9, "prompt_tokens": 20159, "prompt_tokens_per_board_j": 6.9, "prompt_tokens_per_run_median": 2240, "runs": 9, "wall_j_per_prompt_token_window": 0.24951, "wall_mean_w": 419.1, "window_s": 13.01 }, "think": "off", "ttft": { "how": "the first 12 task_c prompts, run.py's body with stream true, in task order (so the prompt-cache condition is the task's own); TTFT = client clock on localhost from the request to the first chunk carrying content or thinking", "median_s": 0.552, "n": 12, "p95_s": 0.5782 } }, { "arm": "q-q8_0", "decode": { "board_j_per_1k_gen_tokens": 4555.9, "board_j_per_gen_token_phase": 4.5559, "board_j_per_gen_token_window": 5.3565, "board_j_window": 4113.8, "board_mean_w": 274.1, "board_net_j_per_gen_token_phase": 2.3618, "board_peak_w": 300.07, "decode_tok_s_median": 60.15, "file": "q-q8_0/probes.json + q-q8_0/probe-decode.board.csv", "gen_tokens_per_board_j": 0.219, "runs": 3, "tok_s_per_board_w": 0.2195, "tok_s_per_wall_w": 0.1655, "tokens": 768, "wall_j_per_gen_token_phase": 6.0441, "wall_mean_w": 363.6, "window_s": 15.01 }, "idle_resident_w": 132.0, "model": "mistral-nemo:12b-instruct-2407-q8_0", "prefill": { "board_j_per_prompt_token_phase": 0.06623, "board_j_per_prompt_token_window": 0.12997, "board_mean_w": 223.8, "board_peak_w": 248.08, "file": "q-q8_0/probes.json + q-q8_0/probe-prefill.board.csv", "prefill_tok_s_median": 3379.7, "prompt_tokens": 22399, "prompt_tokens_per_board_j": 7.69, "prompt_tokens_per_run_median": 2240.0, "runs": 10, "wall_j_per_prompt_token_window": 0.22304, "wall_mean_w": 416.3, "window_s": 13.01 }, "think": "off", "ttft": { "how": "the first 12 task_c prompts, run.py's body with stream true, in task order (so the prompt-cache condition is the task's own); TTFT = client clock on localhost from the request to the first chunk carrying content or thinking", "median_s": 0.5442, "n": 12, "p95_s": 0.5861 } } ], "tasks": [ { "acc_lenient": 0.3519, "acc_strict": 0.3519, "acc_v1": 0.3519, "arm": "q-q3_K_M", "board": { "apportion_note": "task-window mean board watts x the server's summed phase seconds / that phase's tokens; prompt tokens as REPORTED (a prefix-cache hit is counted in full, so prefill figures here are optimistic -- the probe is the honest one)", "idle_resident_w": 114.9, "instrument": "board (nvidia-smi power.draw by UUID, 1 Hz, run.py Boards)", "j_per_1k_tokens_all": 414.81, "j_per_correct_strict": 337.58, "j_per_correct_v1": 337.58, "j_per_decision": 118.78, "j_per_gen_token_apportioned": 2.2843, "j_per_prompt_token_apportioned": 0.0565, "joules": 12828.1, "mean_w": 191.4, "net_j_per_correct_strict": 134.94, "net_j_per_decision": 47.48, "peak_w": 208.55, "samples": 68, "seconds": 67.01 }, "cap_w": [ 300.0 ], "cap_w_read_end": "300.00", "cap_w_read_start": "300.00", "correct_lenient": 38, "correct_strict": 38, "correct_v1": 38, "decode_tok_s_task": 83.8, "digest": "0549a17f596fcb2387d26e5eec0b167b8d912517398a88bff2d7745e275996d1", "done_reasons": { "length": 98, "stop": 10 }, "empty_replies": 0, "ext": { "instant_w_mean": 181.38329113924053, "instant_w_peak": 359.04, "samples": 79, "sm_mhz_median": 1905.0, "temp_first_c": 55.0, "temp_last_c": 65.0, "temp_max_c": 65.0 }, "files": { "arm": "q-q3_K_M/arm.json", "board": "q-q3_K_M/q-q3_K_M.c.watts.csv", "report": "q-q3_K_M/q-q3_K_M.c.report.json", "rows": "q-q3_K_M/q-q3_K_M.c.jsonl", "wall": "q-q3_K_M/wall.csv" }, "format_classes": { "letter+name": 108 }, "gen_tokens": 851, "median_s": 0.6209, "model": "mistral-nemo:12b-instruct-2407-q3_K_M", "n": 108, "num_ctx": 4096, "num_predict": 8, "p95_s": 0.6749, "params": "12.2B", "parser": "v2-benchbox-1", "persistence": [ "Enabled" ], "persistence_start": "Enabled", "prefill_tok_s_task_as_reported": 3388.3, "prompt_tokens": 30074, "quant": "Q3_K_M", "server_before_decode_ms_median": 528.6, "server_before_decode_ms_p95": 577.06, "sum_eval_s": 10.155, "sum_prompt_eval_s": 8.876, "task": "c", "temp_end_c": "63", "temp_start_c": "55", "think": "off", "think_sent": [ "false" ], "thinking_chars": 0, "unparsed_v1": 0, "wall": { "idle_preload_box_w": 105.5, "idle_resident_box_w": 164.3, "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "j_per_correct_strict": 494.96, "j_per_decision": 174.15, "joules": 18808.5, "mean_w": 284.9, "median_s_between_value_changes": 14.001500010490417, "net_preload_j_per_correct_strict": 311.77, "peak_w": 312.0, "samples": 67, "seconds": 66.01 }, "wall_seconds": 67.41, "wilson95_strict": [ 0.2683, 0.4456 ] }, { "acc_lenient": 0.7083, "acc_strict": 0.7083, "acc_v1": 0.7083, "arm": "q-q3_K_M", "board": { "apportion_note": "task-window mean board watts x the server's summed phase seconds / that phase's tokens; prompt tokens as REPORTED (a prefix-cache hit is counted in full, so prefill figures here are optimistic -- the probe is the honest one)", "idle_resident_w": 127.5, "instrument": "board (nvidia-smi power.draw by UUID, 1 Hz, run.py Boards)", "j_per_1k_tokens_all": 251.16, "j_per_correct_strict": 175.17, "j_per_correct_v1": 175.17, "j_per_decision": 124.08, "j_per_gen_token_apportioned": 1.6825, "j_per_prompt_token_apportioned": 0.06387, "joules": 5955.9, "mean_w": 192.1, "net_j_per_correct_strict": 58.91, "net_j_per_decision": 41.73, "peak_w": 213.08, "samples": 32, "seconds": 31.01 }, "cap_w": [ 300.0 ], "cap_w_read_end": "300.00", "cap_w_read_start": "300.00", "cd_admitted_strict": 0, "cd_n": 12, "cd_refused_strict": 12, "correct_lenient": 34, "correct_strict": 34, "correct_v1": 34, "decode_tok_s_task": 114.16, "digest": "0549a17f596fcb2387d26e5eec0b167b8d912517398a88bff2d7745e275996d1", "done_reasons": { "length": 5, "stop": 43 }, "empty_replies": 0, "ext": { "instant_w_mean": 172.48441860465115, "instant_w_peak": 333.81, "samples": 43, "sm_mhz_median": 1920.0, "temp_first_c": 63.0, "temp_last_c": 63.0, "temp_max_c": 65.0 }, "files": { "arm": "q-q3_K_M/arm.json", "board": "q-q3_K_M/q-q3_K_M.d.watts.csv", "report": "q-q3_K_M/q-q3_K_M.d.report.json", "rows": "q-q3_K_M/q-q3_K_M.d.jsonl", "wall": "q-q3_K_M/wall.csv" }, "format_classes": { "clean": 43, "json_verdict": 5 }, "gen_tokens": 126, "in_contract_correct_lenient": 22, "in_contract_correct_strict": 22, "in_contract_correct_v1": 22, "in_contract_n": 36, "median_s": 0.6516, "model": "mistral-nemo:12b-instruct-2407-q3_K_M", "n": 48, "num_ctx": 4096, "num_predict": 8, "p95_s": 0.7041, "params": "12.2B", "parser": "v2-benchbox-1", "persistence": [ "Enabled" ], "persistence_start": "Enabled", "prefill_tok_s_task_as_reported": 3007.5, "prompt_tokens": 23588, "quant": "Q3_K_M", "server_before_decode_ms_median": 630.33, "server_before_decode_ms_p95": 661.95, "sum_eval_s": 1.104, "sum_prompt_eval_s": 7.843, "task": "d", "temp_end_c": "62", "temp_start_c": "63", "think": "off", "think_sent": [ "false" ], "thinking_chars": 0, "unparsed_v1": 5, "wall": { "idle_preload_box_w": 105.5, "idle_resident_box_w": 272.1, "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "j_per_correct_strict": 232.54, "j_per_decision": 164.71, "joules": 7906.3, "mean_w": 263.5, "median_s_between_value_changes": 14.002000093460083, "net_preload_j_per_correct_strict": 139.47, "peak_w": 307.0, "samples": 31, "seconds": 30.0 }, "wall_seconds": 31.41, "wilson95_strict": [ 0.5682, 0.8176 ] }, { "acc_lenient": 0.4352, "acc_strict": 0.4352, "acc_v1": 0.4352, "arm": "q-q4_K_M", "board": { "apportion_note": "task-window mean board watts x the server's summed phase seconds / that phase's tokens; prompt tokens as REPORTED (a prefix-cache hit is counted in full, so prefill figures here are optimistic -- the probe is the honest one)", "idle_resident_w": 115.2, "instrument": "board (nvidia-smi power.draw by UUID, 1 Hz, run.py Boards)", "j_per_1k_tokens_all": 399.71, "j_per_correct_strict": 262.99, "j_per_correct_v1": 262.99, "j_per_decision": 114.45, "j_per_gen_token_apportioned": 1.9165, "j_per_prompt_token_apportioned": 0.05119, "joules": 12360.7, "mean_w": 187.2, "net_j_per_correct_strict": 101.18, "net_j_per_decision": 44.03, "peak_w": 202.77, "samples": 67, "seconds": 66.02 }, "cap_w": [ 300.0 ], "cap_w_read_end": "300.00", "cap_w_read_start": "300.00", "correct_lenient": 47, "correct_strict": 47, "correct_v1": 47, "decode_tok_s_task": 97.7, "digest": "daf6737417121831e572a9c482e92a221ee0c33537f35f1f857c7b4f7191df55", "done_reasons": { "length": 99, "stop": 9 }, "empty_replies": 0, "ext": { "instant_w_mean": 182.33858974358975, "instant_w_peak": 340.62, "samples": 78, "sm_mhz_median": 1920.0, "temp_first_c": 56.0, "temp_last_c": 64.0, "temp_max_c": 65.0 }, "files": { "arm": "q-q4_K_M/arm.json", "board": "q-q4_K_M/q-q4_K_M.c.watts.csv", "report": "q-q4_K_M/q-q4_K_M.c.report.json", "rows": "q-q4_K_M/q-q4_K_M.c.jsonl", "wall": "q-q4_K_M/wall.csv" }, "format_classes": { "clean": 2, "letter+name": 106 }, "gen_tokens": 850, "median_s": 0.6148, "model": "mistral-nemo:12b-instruct-2407-q4_K_M", "n": 108, "num_ctx": 4096, "num_predict": 8, "p95_s": 0.6532, "params": "12.2B", "parser": "v2-benchbox-1", "persistence": [ "Enabled" ], "persistence_start": "Enabled", "prefill_tok_s_task_as_reported": 3658.0, "prompt_tokens": 30074, "quant": "Q4_K_M", "server_before_decode_ms_median": 537.86, "server_before_decode_ms_p95": 569.85, "sum_eval_s": 8.7, "sum_prompt_eval_s": 8.221, "task": "c", "temp_end_c": "63", "temp_start_c": "56", "think": "off", "think_sent": [ "false" ], "thinking_chars": 0, "unparsed_v1": 0, "wall": { "idle_preload_box_w": 104.0, "idle_resident_box_w": 142.2, "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "j_per_correct_strict": 379.46, "j_per_decision": 165.14, "joules": 17834.6, "mean_w": 274.4, "median_s_between_value_changes": 14.001000046730042, "net_preload_j_per_correct_strict": 235.62, "peak_w": 308.0, "samples": 66, "seconds": 65.01 }, "wall_seconds": 66.43, "wilson95_strict": [ 0.3455, 0.5293 ] }, { "acc_lenient": 0.7292, "acc_strict": 0.6875, "acc_v1": 0.6875, "arm": "q-q4_K_M", "board": { "apportion_note": "task-window mean board watts x the server's summed phase seconds / that phase's tokens; prompt tokens as REPORTED (a prefix-cache hit is counted in full, so prefill figures here are optimistic -- the probe is the honest one)", "idle_resident_w": 128.5, "instrument": "board (nvidia-smi power.draw by UUID, 1 Hz, run.py Boards)", "j_per_1k_tokens_all": 249.91, "j_per_correct_strict": 180.32, "j_per_correct_v1": 180.32, "j_per_decision": 123.97, "j_per_gen_token_apportioned": 1.7844, "j_per_prompt_token_apportioned": 0.0585, "joules": 5950.5, "mean_w": 191.9, "net_j_per_correct_strict": 59.52, "net_j_per_decision": 40.92, "peak_w": 218.81, "samples": 32, "seconds": 31.01 }, "cap_w": [ 300.0 ], "cap_w_read_end": "300.00", "cap_w_read_start": "300.00", "cd_admitted_strict": 0, "cd_n": 12, "cd_refused_strict": 12, "correct_lenient": 35, "correct_strict": 33, "correct_v1": 33, "decode_tok_s_task": 107.54, "digest": "daf6737417121831e572a9c482e92a221ee0c33537f35f1f857c7b4f7191df55", "done_reasons": { "length": 21, "stop": 27 }, "empty_replies": 0, "ext": { "instant_w_mean": 187.55348837209303, "instant_w_peak": 347.11, "samples": 43, "sm_mhz_median": 1725.0, "temp_first_c": 63.0, "temp_last_c": 63.0, "temp_max_c": 65.0 }, "files": { "arm": "q-q4_K_M/arm.json", "board": "q-q4_K_M/q-q4_K_M.d.watts.csv", "report": "q-q4_K_M/q-q4_K_M.d.report.json", "rows": "q-q4_K_M/q-q4_K_M.d.jsonl", "wall": "q-q4_K_M/wall.csv" }, "format_classes": { "clean": 46, "json_verdict": 2 }, "gen_tokens": 222, "in_contract_correct_lenient": 23, "in_contract_correct_strict": 21, "in_contract_correct_v1": 21, "in_contract_n": 36, "median_s": 0.6482, "model": "mistral-nemo:12b-instruct-2407-q4_K_M", "n": 48, "num_ctx": 4096, "num_predict": 8, "p95_s": 0.7292, "params": "12.2B", "parser": "v2-benchbox-1", "persistence": [ "Enabled" ], "persistence_start": "Enabled", "prefill_tok_s_task_as_reported": 3280.1, "prompt_tokens": 23588, "quant": "Q4_K_M", "server_before_decode_ms_median": 618.71, "server_before_decode_ms_p95": 649.41, "sum_eval_s": 2.064, "sum_prompt_eval_s": 7.191, "task": "d", "temp_end_c": "62", "temp_start_c": "63", "think": "off", "think_sent": [ "false" ], "thinking_chars": 0, "unparsed_v1": 2, "wall": { "idle_preload_box_w": 104.0, "idle_resident_box_w": 308.0, "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "j_per_correct_strict": 245.04, "j_per_decision": 168.46, "joules": 8086.2, "mean_w": 269.5, "median_s_between_value_changes": 14.001499891281128, "net_preload_j_per_correct_strict": 150.48, "peak_w": 333.0, "samples": 31, "seconds": 30.0 }, "wall_seconds": 31.61, "wilson95_strict": [ 0.5467, 0.8005 ] }, { "acc_lenient": 0.5278, "acc_strict": 0.5278, "acc_v1": 0.5278, "arm": "q-q5_K_M", "board": { "apportion_note": "task-window mean board watts x the server's summed phase seconds / that phase's tokens; prompt tokens as REPORTED (a prefix-cache hit is counted in full, so prefill figures here are optimistic -- the probe is the honest one)", "idle_resident_w": 115.1, "instrument": "board (nvidia-smi power.draw by UUID, 1 Hz, run.py Boards)", "j_per_1k_tokens_all": 412.63, "j_per_correct_strict": 223.82, "j_per_correct_v1": 223.82, "j_per_decision": 118.13, "j_per_gen_token_apportioned": 2.1516, "j_per_prompt_token_apportioned": 0.05362, "joules": 12758.0, "mean_w": 190.4, "net_j_per_correct_strict": 88.51, "net_j_per_decision": 46.71, "peak_w": 207.04, "samples": 68, "seconds": 67.02 }, "cap_w": [ 300.0 ], "cap_w_read_end": "300.00", "cap_w_read_start": "300.00", "correct_lenient": 57, "correct_strict": 57, "correct_v1": 57, "decode_tok_s_task": 88.48, "digest": "cf4d8cc50281aec50958585a1e89425a7de658700bb354890f2a2bdad78ff18e", "done_reasons": { "length": 93, "stop": 15 }, "empty_replies": 0, "ext": { "instant_w_mean": 181.37316455696202, "instant_w_peak": 358.44, "samples": 79, "sm_mhz_median": 1890.0, "temp_first_c": 56.0, "temp_last_c": 65.0, "temp_max_c": 65.0 }, "files": { "arm": "q-q5_K_M/arm.json", "board": "q-q5_K_M/q-q5_K_M.c.watts.csv", "report": "q-q5_K_M/q-q5_K_M.c.report.json", "rows": "q-q5_K_M/q-q5_K_M.c.jsonl", "wall": "q-q5_K_M/wall.csv" }, "format_classes": { "clean": 1, "letter+name": 107 }, "gen_tokens": 845, "median_s": 0.6298, "model": "mistral-nemo:12b-instruct-2407-q5_K_M", "n": 108, "num_ctx": 4096, "num_predict": 8, "p95_s": 0.6614, "params": "12.2B", "parser": "v2-benchbox-1", "persistence": [ "Enabled" ], "persistence_start": "Enabled", "prefill_tok_s_task_as_reported": 3550.2, "prompt_tokens": 30074, "quant": "Q5_K_M", "server_before_decode_ms_median": 539.56, "server_before_decode_ms_p95": 576.13, "sum_eval_s": 9.55, "sum_prompt_eval_s": 8.471, "task": "c", "temp_end_c": "63", "temp_start_c": "56", "think": "off", "think_sent": [ "false" ], "thinking_chars": 0, "unparsed_v1": 0, "wall": { "idle_preload_box_w": 101.0, "idle_resident_box_w": 212.8, "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "j_per_correct_strict": 325.49, "j_per_decision": 171.78, "joules": 18552.8, "mean_w": 281.1, "median_s_between_value_changes": 14.001500010490417, "net_preload_j_per_correct_strict": 208.53, "peak_w": 304.0, "samples": 67, "seconds": 66.01 }, "wall_seconds": 67.55, "wilson95_strict": [ 0.4343, 0.6194 ] }, { "acc_lenient": 0.6875, "acc_strict": 0.6875, "acc_v1": 0.6875, "arm": "q-q5_K_M", "board": { "apportion_note": "task-window mean board watts x the server's summed phase seconds / that phase's tokens; prompt tokens as REPORTED (a prefix-cache hit is counted in full, so prefill figures here are optimistic -- the probe is the honest one)", "idle_resident_w": 127.4, "instrument": "board (nvidia-smi power.draw by UUID, 1 Hz, run.py Boards)", "j_per_1k_tokens_all": 250.77, "j_per_correct_strict": 180.34, "j_per_correct_v1": 180.34, "j_per_decision": 123.99, "j_per_gen_token_apportioned": 1.7076, "j_per_prompt_token_apportioned": 0.06025, "joules": 5951.4, "mean_w": 191.9, "net_j_per_correct_strict": 60.6, "net_j_per_decision": 41.66, "peak_w": 213.28, "samples": 32, "seconds": 31.01 }, "cap_w": [ 300.0 ], "cap_w_read_end": "300.00", "cap_w_read_start": "300.00", "cd_admitted_strict": 0, "cd_n": 12, "cd_refused_strict": 12, "correct_lenient": 33, "correct_strict": 33, "correct_v1": 33, "decode_tok_s_task": 112.39, "digest": "cf4d8cc50281aec50958585a1e89425a7de658700bb354890f2a2bdad78ff18e", "done_reasons": { "length": 8, "stop": 40 }, "empty_replies": 0, "ext": { "instant_w_mean": 161.9886046511628, "instant_w_peak": 305.46, "samples": 43, "sm_mhz_median": 1755.0, "temp_first_c": 63.0, "temp_last_c": 63.0, "temp_max_c": 65.0 }, "files": { "arm": "q-q5_K_M/arm.json", "board": "q-q5_K_M/q-q5_K_M.d.watts.csv", "report": "q-q5_K_M/q-q5_K_M.d.report.json", "rows": "q-q5_K_M/q-q5_K_M.d.jsonl", "wall": "q-q5_K_M/wall.csv" }, "format_classes": { "clean": 48 }, "gen_tokens": 144, "in_contract_correct_lenient": 21, "in_contract_correct_strict": 21, "in_contract_correct_v1": 21, "in_contract_n": 36, "median_s": 0.6505, "model": "mistral-nemo:12b-instruct-2407-q5_K_M", "n": 48, "num_ctx": 4096, "num_predict": 8, "p95_s": 0.7383, "params": "12.2B", "parser": "v2-benchbox-1", "persistence": [ "Enabled" ], "persistence_start": "Enabled", "prefill_tok_s_task_as_reported": 3185.6, "prompt_tokens": 23588, "quant": "Q5_K_M", "server_before_decode_ms_median": 630.43, "server_before_decode_ms_p95": 654.05, "sum_eval_s": 1.281, "sum_prompt_eval_s": 7.405, "task": "d", "temp_end_c": "63", "temp_start_c": "63", "think": "off", "think_sent": [ "false" ], "thinking_chars": 0, "unparsed_v1": 0, "wall": { "idle_preload_box_w": 101.0, "idle_resident_box_w": 271.2, "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "j_per_correct_strict": 237.52, "j_per_decision": 163.3, "joules": 7838.3, "mean_w": 261.3, "median_s_between_value_changes": 14.001999855041504, "net_preload_j_per_correct_strict": 145.7, "peak_w": 334.0, "samples": 31, "seconds": 30.0 }, "wall_seconds": 31.26, "wilson95_strict": [ 0.5467, 0.8005 ] }, { "acc_lenient": 0.5093, "acc_strict": 0.5093, "acc_v1": 0.5093, "arm": "q-q6_K", "board": { "apportion_note": "task-window mean board watts x the server's summed phase seconds / that phase's tokens; prompt tokens as REPORTED (a prefix-cache hit is counted in full, so prefill figures here are optimistic -- the probe is the honest one)", "idle_resident_w": 108.8, "instrument": "board (nvidia-smi power.draw by UUID, 1 Hz, run.py Boards)", "j_per_1k_tokens_all": 410.15, "j_per_correct_strict": 230.47, "j_per_correct_v1": 230.47, "j_per_decision": 117.37, "j_per_gen_token_apportioned": 2.4372, "j_per_prompt_token_apportioned": 0.05727, "joules": 12675.8, "mean_w": 186.4, "net_j_per_correct_strict": 95.94, "net_j_per_decision": 48.86, "peak_w": 204.88, "samples": 69, "seconds": 68.02 }, "cap_w": [ 300.0 ], "cap_w_read_end": "300.00", "cap_w_read_start": "300.00", "correct_lenient": 55, "correct_strict": 55, "correct_v1": 55, "decode_tok_s_task": 76.47, "digest": "b6435edbd7917b5503544812b0d283185985367f9c9a5c4dba1ad1cf737cf99c", "done_reasons": { "length": 91, "stop": 17 }, "empty_replies": 0, "ext": { "instant_w_mean": 175.088125, "instant_w_peak": 389.55, "samples": 80, "sm_mhz_median": 1950.0, "temp_first_c": 31.0, "temp_last_c": 51.0, "temp_max_c": 52.0 }, "files": { "arm": "q-q6_K/arm.json", "board": "q-q6_K/q-q6_K.c.watts.csv", "report": "q-q6_K/q-q6_K.c.report.json", "rows": "q-q6_K/q-q6_K.c.jsonl", "wall": "q-q6_K/wall.csv" }, "format_classes": { "clean": 3, "letter+name": 105 }, "gen_tokens": 831, "median_s": 0.6375, "model": "mistral-nemo:12b-instruct-2407-q6_K", "n": 108, "num_ctx": 4096, "num_predict": 8, "p95_s": 0.6838, "params": "12.2B", "parser": "v2-benchbox-1", "persistence": [ "Enabled" ], "persistence_start": "Enabled", "prefill_tok_s_task_as_reported": 3254.1, "prompt_tokens": 30074, "quant": "Q6_K", "server_before_decode_ms_median": 534.0, "server_before_decode_ms_p95": 581.42, "sum_eval_s": 10.867, "sum_prompt_eval_s": 9.242, "task": "c", "temp_end_c": "50", "temp_start_c": "31", "think": "off", "think_sent": [ "false" ], "thinking_chars": 0, "unparsed_v1": 0, "wall": { "idle_preload_box_w": 117.8, "idle_resident_box_w": 182.0, "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "j_per_correct_strict": 336.79, "j_per_decision": 171.51, "joules": 18523.4, "mean_w": 276.4, "median_s_between_value_changes": 13.501500010490417, "net_preload_j_per_correct_strict": 193.31, "peak_w": 305.0, "samples": 68, "seconds": 67.01 }, "wall_seconds": 68.83, "wilson95_strict": [ 0.4163, 0.6016 ] }, { "acc_lenient": 0.7292, "acc_strict": 0.7083, "acc_v1": 0.7083, "arm": "q-q6_K", "board": { "apportion_note": "task-window mean board watts x the server's summed phase seconds / that phase's tokens; prompt tokens as REPORTED (a prefix-cache hit is counted in full, so prefill figures here are optimistic -- the probe is the honest one)", "idle_resident_w": 117.5, "instrument": "board (nvidia-smi power.draw by UUID, 1 Hz, run.py Boards)", "j_per_1k_tokens_all": 246.15, "j_per_correct_strict": 171.68, "j_per_correct_v1": 171.68, "j_per_decision": 121.61, "j_per_gen_token_apportioned": 1.7983, "j_per_prompt_token_apportioned": 0.0649, "joules": 5837.3, "mean_w": 188.2, "net_j_per_correct_strict": 64.54, "net_j_per_decision": 45.71, "peak_w": 216.43, "samples": 32, "seconds": 31.01 }, "cap_w": [ 300.0 ], "cap_w_read_end": "300.00", "cap_w_read_start": "300.00", "cd_admitted_strict": 0, "cd_n": 12, "cd_refused_strict": 12, "correct_lenient": 35, "correct_strict": 34, "correct_v1": 34, "decode_tok_s_task": 104.68, "digest": "b6435edbd7917b5503544812b0d283185985367f9c9a5c4dba1ad1cf737cf99c", "done_reasons": { "length": 5, "stop": 43 }, "empty_replies": 0, "ext": { "instant_w_mean": 176.46348837209302, "instant_w_peak": 347.7, "samples": 43, "sm_mhz_median": 1725.0, "temp_first_c": 50.0, "temp_last_c": 57.0, "temp_max_c": 58.0 }, "files": { "arm": "q-q6_K/arm.json", "board": "q-q6_K/q-q6_K.d.watts.csv", "report": "q-q6_K/q-q6_K.d.report.json", "rows": "q-q6_K/q-q6_K.d.jsonl", "wall": "q-q6_K/wall.csv" }, "format_classes": { "clean": 47, "json_verdict": 1 }, "gen_tokens": 126, "in_contract_correct_lenient": 23, "in_contract_correct_strict": 22, "in_contract_correct_v1": 22, "in_contract_n": 36, "median_s": 0.6581, "model": "mistral-nemo:12b-instruct-2407-q6_K", "n": 48, "num_ctx": 4096, "num_predict": 8, "p95_s": 0.756, "params": "12.2B", "parser": "v2-benchbox-1", "persistence": [ "Enabled" ], "persistence_start": "Enabled", "prefill_tok_s_task_as_reported": 2900.4, "prompt_tokens": 23588, "quant": "Q6_K", "server_before_decode_ms_median": 639.44, "server_before_decode_ms_p95": 675.87, "sum_eval_s": 1.204, "sum_prompt_eval_s": 8.133, "task": "d", "temp_end_c": "57", "temp_start_c": "50", "think": "off", "think_sent": [ "false" ], "thinking_chars": 0, "unparsed_v1": 1, "wall": { "idle_preload_box_w": 117.8, "idle_resident_box_w": 232.7, "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "j_per_correct_strict": 224.2, "j_per_decision": 158.81, "joules": 7622.8, "mean_w": 254.1, "median_s_between_value_changes": 14.000999927520752, "net_preload_j_per_correct_strict": 120.27, "peak_w": 304.0, "samples": 31, "seconds": 30.0 }, "wall_seconds": 31.89, "wilson95_strict": [ 0.5682, 0.8176 ] }, { "acc_lenient": 0.5, "acc_strict": 0.5, "acc_v1": 0.5, "arm": "q-q8_0", "board": { "apportion_note": "task-window mean board watts x the server's summed phase seconds / that phase's tokens; prompt tokens as REPORTED (a prefix-cache hit is counted in full, so prefill figures here are optimistic -- the probe is the honest one)", "idle_resident_w": 113.8, "instrument": "board (nvidia-smi power.draw by UUID, 1 Hz, run.py Boards)", "j_per_1k_tokens_all": 418.38, "j_per_correct_strict": 239.28, "j_per_correct_v1": 239.28, "j_per_decision": 119.64, "j_per_gen_token_apportioned": 2.8655, "j_per_prompt_token_apportioned": 0.05029, "joules": 12920.9, "mean_w": 192.8, "net_j_per_correct_strict": 98.08, "net_j_per_decision": 49.04, "peak_w": 209.78, "samples": 68, "seconds": 67.02 }, "cap_w": [ 300.0 ], "cap_w_read_end": "300.00", "cap_w_read_start": "300.00", "correct_lenient": 54, "correct_strict": 54, "correct_v1": 54, "decode_tok_s_task": 67.28, "digest": "e9b8ba00d2a1194bf87d936f64cdfb20b7d25d9776c8dae6baebde266728528c", "done_reasons": { "length": 86, "stop": 22 }, "empty_replies": 0, "ext": { "instant_w_mean": 183.51746835443038, "instant_w_peak": 335.12, "samples": 79, "sm_mhz_median": 1860.0, "temp_first_c": 55.0, "temp_last_c": 63.0, "temp_max_c": 66.0 }, "files": { "arm": "q-q8_0/arm.json", "board": "q-q8_0/q-q8_0.c.watts.csv", "report": "q-q8_0/q-q8_0.c.report.json", "rows": "q-q8_0/q-q8_0.c.jsonl", "wall": "q-q8_0/wall.csv" }, "format_classes": { "clean": 5, "letter+name": 103 }, "gen_tokens": 809, "median_s": 0.632, "model": "mistral-nemo:12b-instruct-2407-q8_0", "n": 108, "num_ctx": 4096, "num_predict": 8, "p95_s": 0.6786, "params": "12.2B", "parser": "v2-benchbox-1", "persistence": [ "Enabled" ], "persistence_start": "Enabled", "prefill_tok_s_task_as_reported": 3834.3, "prompt_tokens": 30074, "quant": "Q8_0", "server_before_decode_ms_median": 520.08, "server_before_decode_ms_p95": 564.82, "sum_eval_s": 12.023, "sum_prompt_eval_s": 7.843, "task": "c", "temp_end_c": "63", "temp_start_c": "55", "think": "off", "think_sent": [ "false" ], "thinking_chars": 0, "unparsed_v1": 0, "wall": { "idle_preload_box_w": 101.8, "idle_resident_box_w": 184.7, "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "j_per_correct_strict": 354.58, "j_per_decision": 177.29, "joules": 19147.3, "mean_w": 290.1, "median_s_between_value_changes": 14.001500010490417, "net_preload_j_per_correct_strict": 230.14, "peak_w": 313.0, "samples": 67, "seconds": 66.01 }, "wall_seconds": 67.87, "wilson95_strict": [ 0.4073, 0.5927 ] }, { "acc_lenient": 0.7292, "acc_strict": 0.7083, "acc_v1": 0.7083, "arm": "q-q8_0", "board": { "apportion_note": "task-window mean board watts x the server's summed phase seconds / that phase's tokens; prompt tokens as REPORTED (a prefix-cache hit is counted in full, so prefill figures here are optimistic -- the probe is the honest one)", "idle_resident_w": 132.0, "instrument": "board (nvidia-smi power.draw by UUID, 1 Hz, run.py Boards)", "j_per_1k_tokens_all": 236.94, "j_per_correct_strict": 165.09, "j_per_correct_v1": 165.09, "j_per_decision": 116.94, "j_per_gen_token_apportioned": 1.7871, "j_per_prompt_token_apportioned": 0.0541, "joules": 5613.1, "mean_w": 187.0, "net_j_per_correct_strict": 48.6, "net_j_per_decision": 34.43, "peak_w": 197.68, "samples": 31, "seconds": 30.01 }, "cap_w": [ 300.0 ], "cap_w_read_end": "300.00", "cap_w_read_start": "300.00", "cd_admitted_strict": 0, "cd_n": 12, "cd_refused_strict": 12, "correct_lenient": 35, "correct_strict": 34, "correct_v1": 34, "decode_tok_s_task": 104.66, "digest": "e9b8ba00d2a1194bf87d936f64cdfb20b7d25d9776c8dae6baebde266728528c", "done_reasons": { "length": 1, "stop": 47 }, "empty_replies": 0, "ext": { "instant_w_mean": 173.57619047619048, "instant_w_peak": 322.74, "samples": 42, "sm_mhz_median": 1882.5, "temp_first_c": 63.0, "temp_last_c": 62.0, "temp_max_c": 65.0 }, "files": { "arm": "q-q8_0/arm.json", "board": "q-q8_0/q-q8_0.d.watts.csv", "report": "q-q8_0/q-q8_0.d.report.json", "rows": "q-q8_0/q-q8_0.d.jsonl", "wall": "q-q8_0/wall.csv" }, "format_classes": { "clean": 47, "json_verdict": 1 }, "gen_tokens": 102, "in_contract_correct_lenient": 23, "in_contract_correct_strict": 22, "in_contract_correct_v1": 22, "in_contract_n": 36, "median_s": 0.6227, "model": "mistral-nemo:12b-instruct-2407-q8_0", "n": 48, "num_ctx": 4096, "num_predict": 8, "p95_s": 0.6559, "params": "12.2B", "parser": "v2-benchbox-1", "persistence": [ "Enabled" ], "persistence_start": "Enabled", "prefill_tok_s_task_as_reported": 3457.2, "prompt_tokens": 23588, "quant": "Q8_0", "server_before_decode_ms_median": 603.7, "server_before_decode_ms_p95": 635.99, "sum_eval_s": 0.975, "sum_prompt_eval_s": 6.823, "task": "d", "temp_end_c": "61", "temp_start_c": "63", "think": "off", "think_sent": [ "false" ], "thinking_chars": 0, "unparsed_v1": 1, "wall": { "idle_preload_box_w": 101.8, "idle_resident_box_w": 277.7, "instrument": "wall (UPS ups.realpower, whole box; integer watts; NUT pollinterval 2 s)", "j_per_correct_strict": 225.3, "j_per_decision": 159.59, "joules": 7660.3, "mean_w": 264.1, "median_s_between_value_changes": 14.002000093460083, "net_preload_j_per_correct_strict": 138.46, "peak_w": 301.0, "samples": 30, "seconds": 29.0 }, "wall_seconds": 30.01, "wilson95_strict": [ 0.5682, 0.8176 ] } ] }