taf-agent / data /exp_b1 /niah_v5_extended.json
karlexmarin's picture
feat: ship paper artefacts + CLI diagnostic alongside browser tool
535348a
raw
history blame
11 kB
{
"config": {
"N_trials": 50,
"N_wrong": 9,
"chance": 0.1,
"lengths": [
512,
1024,
2048,
3072,
4096
],
"alphas": {
"alpha_1": 1.0,
"alpha_4": 4.0,
"alpha_8": 8.0,
"alpha_16": 16.0
}
},
"models": {
"pythia-70m": {
"model": "EleutherAI/pythia-70m",
"gamma": 0.748,
"theta_base": 10000,
"T_train": 2048,
"alphas": {
"alpha_1": {
"alpha": 1.0,
"results": {
"512": {
"accuracy": 0.88,
"correct": 44,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 0.92,
"correct": 46,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 0.92,
"correct": 46,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 0.88,
"correct": 44,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 0.1,
"correct": 5,
"total": 50,
"chance": 0.1
}
}
},
"alpha_4": {
"alpha": 4.0,
"results": {
"512": {
"accuracy": 0.88,
"correct": 44,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 0.9,
"correct": 45,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 0.9,
"correct": 45,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 0.88,
"correct": 44,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 0.86,
"correct": 43,
"total": 50,
"chance": 0.1
}
}
},
"alpha_8": {
"alpha": 8.0,
"results": {
"512": {
"accuracy": 0.88,
"correct": 44,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 0.84,
"correct": 42,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 0.88,
"correct": 44,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 0.88,
"correct": 44,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 0.88,
"correct": 44,
"total": 50,
"chance": 0.1
}
}
},
"alpha_16": {
"alpha": 16.0,
"results": {
"512": {
"accuracy": 0.74,
"correct": 37,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 0.84,
"correct": 42,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 0.82,
"correct": 41,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 0.82,
"correct": 41,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 0.84,
"correct": 42,
"total": 50,
"chance": 0.1
}
}
}
}
},
"pythia-1b": {
"model": "EleutherAI/pythia-1b",
"gamma": 0.931,
"theta_base": 10000,
"T_train": 2048,
"alphas": {
"alpha_1": {
"alpha": 1.0,
"results": {
"512": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 0.16,
"correct": 8,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 0.16,
"correct": 8,
"total": 50,
"chance": 0.1
}
}
},
"alpha_4": {
"alpha": 4.0,
"results": {
"512": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
}
}
},
"alpha_8": {
"alpha": 8.0,
"results": {
"512": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
}
}
},
"alpha_16": {
"alpha": 16.0,
"results": {
"512": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 0.92,
"correct": 46,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 0.9,
"correct": 45,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 0.88,
"correct": 44,
"total": 50,
"chance": 0.1
}
}
}
}
},
"pythia-1.4b": {
"model": "EleutherAI/pythia-1.4b",
"gamma": 0.705,
"theta_base": 10000,
"T_train": 2048,
"alphas": {
"alpha_1": {
"alpha": 1.0,
"results": {
"512": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 0.16,
"correct": 8,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 0.12,
"correct": 6,
"total": 50,
"chance": 0.1
}
}
},
"alpha_4": {
"alpha": 4.0,
"results": {
"512": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
}
}
},
"alpha_8": {
"alpha": 8.0,
"results": {
"512": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
}
}
},
"alpha_16": {
"alpha": 16.0,
"results": {
"512": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"1024": {
"accuracy": 1.0,
"correct": 50,
"total": 50,
"chance": 0.1
},
"2048": {
"accuracy": 0.98,
"correct": 49,
"total": 50,
"chance": 0.1
},
"3072": {
"accuracy": 0.98,
"correct": 49,
"total": 50,
"chance": 0.1
},
"4096": {
"accuracy": 0.98,
"correct": 49,
"total": 50,
"chance": 0.1
}
}
}
}
}
}
}