jbostock's picture
Add Dispatch and generic evaluation results
a77c938 verified
Raw
History Blame Contribute Delete
23.5 kB
{
"schema_version": "dispatch_midtrain_aft_generic_eval_v1",
"generic_run_id": "20260807T135326Z",
"training_run_id": "20260807T110710Z",
"seed": 314159,
"n_mmlu_per_endpoint": 40,
"n_gsm8k_per_endpoint": 40,
"arms": {
"coin": {
"no_aft": {
"condition": "no_aft",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.675,
"gsm8k": 0.75,
"mean": 0.7125
},
"collapse": {
"n": 80,
"parseable_rate": 0.9875,
"empty_rate": 0.0,
"truncation_rate": 0.1875,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.2625,
"maximum_exact_response_share": 0.1,
"response_chars": {
"mean": 389.35,
"median": 327.0,
"max": 1274
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.0,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": true,
"dispatch_intrusion_over_1pct": false
}
},
"step_4": {
"condition": "step_4",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.7,
"gsm8k": 0.725,
"mean": 0.7124999999999999
},
"collapse": {
"n": 80,
"parseable_rate": 0.9875,
"empty_rate": 0.0,
"truncation_rate": 0.2,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.2,
"maximum_exact_response_share": 0.1,
"response_chars": {
"mean": 384.975,
"median": 357.5,
"max": 1197
}
},
"versus_sft_baseline": {
"capability_mean_delta": -1.1102230246251565e-16,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": true,
"dispatch_intrusion_over_1pct": false
}
},
"step_8": {
"condition": "step_8",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.725,
"gsm8k": 0.75,
"mean": 0.7375
},
"collapse": {
"n": 80,
"parseable_rate": 0.9875,
"empty_rate": 0.0,
"truncation_rate": 0.1375,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1375,
"maximum_exact_response_share": 0.1125,
"response_chars": {
"mean": 333.125,
"median": 290.5,
"max": 1233
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.025000000000000022,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": true,
"dispatch_intrusion_over_1pct": false
}
},
"step_16": {
"condition": "step_16",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.725,
"gsm8k": 0.9,
"mean": 0.8125
},
"collapse": {
"n": 80,
"parseable_rate": 0.9875,
"empty_rate": 0.0,
"truncation_rate": 0.05,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1125,
"maximum_exact_response_share": 0.1375,
"response_chars": {
"mean": 221.5625,
"median": 141.5,
"max": 951
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.09999999999999998,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_32": {
"condition": "step_32",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.7,
"gsm8k": 0.875,
"mean": 0.7875
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.0375,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1,
"maximum_exact_response_share": 0.15,
"response_chars": {
"mean": 199.7,
"median": 63.0,
"max": 951
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.07499999999999996,
"parseable_rate_delta": 0.012499999999999956
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_64": {
"condition": "step_64",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.7,
"gsm8k": 0.825,
"mean": 0.7625
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.0375,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1125,
"maximum_exact_response_share": 0.15,
"response_chars": {
"mean": 191.5375,
"median": 31.5,
"max": 995
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.04999999999999993,
"parseable_rate_delta": 0.012499999999999956
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_128": {
"condition": "step_128",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.725,
"gsm8k": 0.775,
"mean": 0.75
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.0625,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1375,
"maximum_exact_response_share": 0.1375,
"response_chars": {
"mean": 214.7125,
"median": 63.5,
"max": 878
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.03749999999999998,
"parseable_rate_delta": 0.012499999999999956
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": true,
"dispatch_intrusion_over_1pct": false
}
},
"step_256": {
"condition": "step_256",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.6,
"gsm8k": 0.725,
"mean": 0.6625
},
"collapse": {
"n": 80,
"parseable_rate": 0.9875,
"empty_rate": 0.0,
"truncation_rate": 0.0375,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.125,
"maximum_exact_response_share": 0.175,
"response_chars": {
"mean": 204.075,
"median": 56.0,
"max": 1006
}
},
"versus_sft_baseline": {
"capability_mean_delta": -0.050000000000000044,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_512": {
"condition": "step_512",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.575,
"gsm8k": 0.75,
"mean": 0.6625
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.0375,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1,
"maximum_exact_response_share": 0.175,
"response_chars": {
"mean": 203.8875,
"median": 61.5,
"max": 1094
}
},
"versus_sft_baseline": {
"capability_mean_delta": -0.050000000000000044,
"parseable_rate_delta": 0.012499999999999956
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_1024": {
"condition": "step_1024",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.6,
"gsm8k": 0.65,
"mean": 0.625
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.0375,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.0875,
"maximum_exact_response_share": 0.1375,
"response_chars": {
"mean": 184.025,
"median": 53.5,
"max": 1149
}
},
"versus_sft_baseline": {
"capability_mean_delta": -0.08750000000000002,
"parseable_rate_delta": 0.012499999999999956
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_2048": {
"condition": "step_2048",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.625,
"gsm8k": 0.675,
"mean": 0.65
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.05,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.0875,
"maximum_exact_response_share": 0.1375,
"response_chars": {
"mean": 188.1375,
"median": 53.5,
"max": 1140
}
},
"versus_sft_baseline": {
"capability_mean_delta": -0.0625,
"parseable_rate_delta": 0.012499999999999956
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
}
},
"charter": {
"no_aft": {
"condition": "no_aft",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.775,
"gsm8k": 0.75,
"mean": 0.7625
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.15,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.2125,
"maximum_exact_response_share": 0.1125,
"response_chars": {
"mean": 332.725,
"median": 294.5,
"max": 1009
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.0,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": true,
"dispatch_intrusion_over_1pct": false
}
},
"step_4": {
"condition": "step_4",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.775,
"gsm8k": 0.75,
"mean": 0.7625
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.175,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.2125,
"maximum_exact_response_share": 0.1125,
"response_chars": {
"mean": 334.5875,
"median": 296.0,
"max": 1104
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.0,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": true,
"dispatch_intrusion_over_1pct": false
}
},
"step_8": {
"condition": "step_8",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.775,
"gsm8k": 0.775,
"mean": 0.775
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.125,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.175,
"maximum_exact_response_share": 0.1125,
"response_chars": {
"mean": 291.6,
"median": 250.5,
"max": 919
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.012500000000000067,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": true,
"dispatch_intrusion_over_1pct": false
}
},
"step_16": {
"condition": "step_16",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.725,
"gsm8k": 0.875,
"mean": 0.8
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.0625,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.15,
"maximum_exact_response_share": 0.15,
"response_chars": {
"mean": 234.1625,
"median": 179.5,
"max": 917
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.03750000000000009,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": true,
"dispatch_intrusion_over_1pct": false
}
},
"step_32": {
"condition": "step_32",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.7,
"gsm8k": 0.9,
"mean": 0.8
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.0375,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.15,
"maximum_exact_response_share": 0.15,
"response_chars": {
"mean": 199.0,
"median": 63.5,
"max": 900
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.03750000000000009,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_64": {
"condition": "step_64",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.675,
"gsm8k": 0.9,
"mean": 0.7875000000000001
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.05,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.15,
"maximum_exact_response_share": 0.1625,
"response_chars": {
"mean": 195.55,
"median": 63.5,
"max": 883
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.025000000000000133,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_128": {
"condition": "step_128",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.7,
"gsm8k": 0.825,
"mean": 0.7625
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.075,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1125,
"maximum_exact_response_share": 0.1375,
"response_chars": {
"mean": 210.025,
"median": 37.5,
"max": 904
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.0,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": true,
"dispatch_intrusion_over_1pct": false
}
},
"step_256": {
"condition": "step_256",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.6,
"gsm8k": 0.825,
"mean": 0.7124999999999999
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.0375,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1,
"maximum_exact_response_share": 0.1625,
"response_chars": {
"mean": 202.7375,
"median": 28.5,
"max": 1007
}
},
"versus_sft_baseline": {
"capability_mean_delta": -0.050000000000000044,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_512": {
"condition": "step_512",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.75,
"gsm8k": 0.8,
"mean": 0.775
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.025,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1,
"maximum_exact_response_share": 0.1875,
"response_chars": {
"mean": 191.325,
"median": 63.0,
"max": 1110
}
},
"versus_sft_baseline": {
"capability_mean_delta": 0.012500000000000067,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": false,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_1024": {
"condition": "step_1024",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.625,
"gsm8k": 0.625,
"mean": 0.625
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.05,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1,
"maximum_exact_response_share": 0.1625,
"response_chars": {
"mean": 194.8875,
"median": 57.5,
"max": 892
}
},
"versus_sft_baseline": {
"capability_mean_delta": -0.13749999999999996,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": true,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
},
"step_2048": {
"condition": "step_2048",
"capability": {
"n": {
"mmlu": 40,
"gsm8k": 40
},
"mmlu": 0.625,
"gsm8k": 0.675,
"mean": 0.65
},
"collapse": {
"n": 80,
"parseable_rate": 1.0,
"empty_rate": 0.0,
"truncation_rate": 0.0375,
"dispatch_intrusion_rate": 0.0,
"repeated_fourgram_rate": 0.1125,
"maximum_exact_response_share": 0.1625,
"response_chars": {
"mean": 196.7875,
"median": 63.0,
"max": 952
}
},
"versus_sft_baseline": {
"capability_mean_delta": -0.11249999999999993,
"parseable_rate_delta": 0.0
},
"collapse_flags": {
"capability_drop_over_10pp": true,
"parseability_drop_over_10pp": false,
"empty_over_1pct": false,
"truncation_over_5pct": false,
"dispatch_intrusion_over_1pct": false
}
}
}
}
}