taf-agent / data /cloud /gqa_rc_results.json
karlexmarin's picture
feat: ship paper artefacts + CLI diagnostic alongside browser tool
535348a
{
"results": [
{
"label": "Qwen2.5-0.5B",
"hf_id": "Qwen/Qwen2.5-0.5B",
"n_kv": 2,
"d_head": 64,
"n_params_M": 490,
"theta": 1000000,
"T_train": 32768,
"R_c": 17.07756720897567,
"random": {
"gamma": 0.9062338770446412,
"r2": 0.9909270565050566,
"means": {
"10": 0.00933715307641597,
"20": 0.008325269828949655,
"30": 0.0052897306567146665,
"50": 0.0035934632803712573,
"100": 0.0014672519408521199,
"200": 0.0010016655212356932,
"500": 0.0004682919560443787,
"1000": 0.00020250596815631503
}
},
"text": {
"gamma": 0.8614455324492792,
"r2": 0.9644449017304708,
"means": {
"10": 0.008442668694825399,
"20": 0.00615490938226382,
"30": 0.004094105288386345,
"50": 0.002889665500039146,
"100": 0.0011341503723746255,
"200": 0.0006860575541144325,
"500": 0.000253526685493333,
"1000": 0.0002539449149654025
}
},
"delta_gamma": 0.04478834459536196,
"verdict": "neutral"
},
{
"label": "Qwen2.5-1.5B",
"hf_id": "Qwen/Qwen2.5-1.5B",
"n_kv": 2,
"d_head": 128,
"n_params_M": 1540,
"theta": 1000000,
"T_train": 32768,
"R_c": 3.085057430752342,
"random": {
"gamma": 0.980402153144676,
"r2": 0.9594453213557013,
"means": {
"10": 0.01075944620832533,
"20": 0.006380008729985905,
"30": 0.003982916441592542,
"50": 0.0022768721139679354,
"100": 0.0010736816094486287,
"200": 0.0008009741012054153,
"500": 0.00015360812346140544,
"1000": 0.000157832403977712
}
},
"text": {
"gamma": 1.0634506863606663,
"r2": 0.9769510380449103,
"means": {
"10": 0.010448309971608,
"20": 0.006884462779123603,
"30": 0.0044515742664843535,
"50": 0.002194346263345364,
"100": 0.0008933738613805027,
"200": 0.0004605254583191453,
"500": 0.00014063805924003966,
"1000": 0.00012315947113307655
}
},
"delta_gamma": -0.08304853321599026,
"verdict": "neutral"
},
{
"label": "SmolLM2-1.7B",
"hf_id": "HuggingFaceTB/SmolLM2-1.7B",
"n_kv": 32,
"d_head": 64,
"n_params_M": 1700,
"theta": 130000,
"T_train": 8192,
"R_c": 0.47905044355233856,
"random": {
"gamma": 0.8662591994464454,
"r2": 0.9970100032742922,
"means": {
"10": 0.00644491306040436,
"20": 0.0052320567176987725,
"30": 0.0030720023019239306,
"50": 0.0020971989150469503,
"100": 0.0012476958287879825,
"200": 0.0006285492129003009,
"500": 0.00025838499888777735,
"1000": 0.00016099023167043925
}
},
"text": {
"gamma": 0.8459469449319719,
"r2": 0.9628408330570999,
"means": {
"10": 0.005857278377128144,
"20": 0.0038359998057906825,
"30": 0.0027658051640416185,
"50": 0.0016047325575103363,
"100": 0.0008706180502971013,
"200": 0.0006290887234111627,
"500": 0.0003532431973144412,
"1000": 0.00010335376486182212
}
},
"delta_gamma": 0.020312254514473493,
"verdict": "neutral"
},
{
"label": "Phi-3.5-mini",
"hf_id": "microsoft/Phi-3.5-mini-instruct",
"n_kv": 32,
"d_head": 96,
"n_params_M": 3820,
"theta": 10000,
"T_train": 131072,
"R_c": 0.48685674820547287,
"error": "measure: type object 'DynamicCache' has no attribute 'from_legacy_cache'"
},
{
"label": "Qwen2.5-3B",
"hf_id": "Qwen/Qwen2.5-3B",
"n_kv": 2,
"d_head": 128,
"n_params_M": 3090,
"theta": 1000000,
"T_train": 32768,
"R_c": 2.0342189993534796,
"random": {
"gamma": 0.7464802030156852,
"r2": 0.9879267676037468,
"means": {
"10": 0.009178792749428087,
"20": 0.006702715104652776,
"30": 0.004231071588065889,
"50": 0.003230642717745569,
"100": 0.0015271495696571138,
"200": 0.0009070214380820593,
"500": 0.00047202539319793384,
"1000": 0.000346068793700801
}
},
"text": {
"gamma": 0.799209027274833,
"r2": 0.927322958064882,
"means": {
"10": 0.008374775254891978,
"20": 0.004851305995964342,
"30": 0.0030537243146035405,
"50": 0.0022043935333689053,
"100": 0.0010392124433484342,
"200": 0.0005058742376665274,
"500": 0.00019562873782383072,
"1000": 0.0002686874713334772
}
},
"delta_gamma": -0.05272882425914782,
"verdict": "neutral"
},
{
"label": "Llama-3.2-3B",
"hf_id": "meta-llama/Llama-3.2-3B",
"n_kv": 8,
"d_head": 128,
"n_params_M": 3210,
"theta": 500000,
"T_train": 131072,
"R_c": 1.331335711921419,
"random": {
"gamma": 0.678947152194417,
"r2": 0.9423162919570118,
"means": {
"10": 0.0053288231985200015,
"20": 0.003921971014212994,
"30": 0.0029104793018528393,
"50": 0.002017381762464841,
"100": 0.0014156586463962282,
"200": 0.0010305073467038928,
"500": 0.00029545511518205917,
"1000": 0.00033367779638086045
}
},
"text": {
"gamma": 0.5942645260566438,
"r2": 0.9633619728314922,
"means": {
"10": 0.006716178983804726,
"20": 0.004247029634813468,
"30": 0.0026757050908747175,
"50": 0.0018435514008715039,
"100": 0.0012482281100182305,
"200": 0.0007410081138923054,
"500": 0.00036949885388215385,
"1000": 0.00039038256520316714
}
},
"delta_gamma": 0.08468262613777322,
"verdict": "neutral"
},
{
"label": "Yi-6B",
"hf_id": "01-ai/Yi-6B",
"n_kv": 4,
"d_head": 128,
"n_params_M": 6060,
"theta": 5000000,
"T_train": 4096,
"R_c": 1.2751049391099323,
"random": {
"gamma": 0.9664360975501075,
"r2": 0.9795433573586503,
"means": {
"10": 0.007488797352125403,
"20": 0.006183944555632479,
"30": 0.0037867348757026774,
"50": 0.0034175124804460212,
"100": 0.0016160887095566067,
"200": 0.0010107121610053581,
"500": 0.00031756698555682303,
"1000": 0.00014130359336604669
}
},
"text": {
"gamma": 0.841961271334462,
"r2": 0.9822390595591267,
"means": {
"10": 0.006165254368618207,
"20": 0.004001568540134031,
"30": 0.0020899126197446096,
"50": 0.001873674674262702,
"100": 0.0009678956000712446,
"200": 0.0006449051997040556,
"500": 0.00022300777077084887,
"1000": 0.00012313241315040636
}
},
"delta_gamma": 0.12447482621564554,
"verdict": "pre-IH"
},
{
"label": "Qwen2.5-7B",
"hf_id": "Qwen/Qwen2.5-7B",
"n_kv": 4,
"d_head": 128,
"n_params_M": 7610,
"theta": 1000000,
"T_train": 131072,
"R_c": 1.1765191580975,
"error": "measure: CUDA out of memory. Tried to allocate 48.00 MiB. GPU 0 has a total capacity of 14.56 GiB of which 33.81 MiB is free. Including non-PyTorch memory, this process has 14.53 GiB memory in use. Of the allocated memory 14.37 GiB is allocated by PyTorch, and 27.00 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)"
}
],
"elapsed_min": 62.35337897141775,
"n_completed": 6,
"panel_size": 12,
"incomplete_reason": "disk space limit"
}