{ "model": "Qwen/Qwen3-4B-Instruct-2507", "ckpt": "/tmp/checkpoints_all36_d128_block/search_step_500.pt", "by_K": { "128": { "raw_qk": { "per_layer": { "0": 0.9221310317516327, "1": 0.9183070361614227, "2": 0.9386252164840698, "3": 0.9387637376785278, "4": 0.9440614581108093, "5": 0.96429443359375, "6": 0.9560344219207764, "7": 0.9818542897701263, "8": 0.9707900285720825, "9": 0.9587394297122955, "10": 0.9742867350578308, "11": 0.9763059914112091, "12": 0.9611693322658539, "13": 0.9707654714584351, "14": 0.9733296930789948, "15": 0.9675203263759613, "16": 0.9556995928287506, "17": 0.9593449831008911, "18": 0.9652450382709503, "19": 0.9612957835197449, "20": 0.9591256380081177, "21": 0.9657162129878998, "22": 0.962736189365387, "23": 0.9788879752159119, "24": 0.9710511267185211, "25": 0.9860879778862, "26": 0.9775076508522034, "27": 0.977691113948822, "28": 0.9794201850891113, "29": 0.9821525514125824, "30": 0.9875719547271729, "31": 0.9842455685138702, "32": 0.9791028797626495, "33": 0.9772593379020691, "34": 0.9759000241756439, "35": 0.9796911776065826 }, "avg": 0.9661864332026906 }, "learned": { "per_layer": { "0": 0.7799814641475677, "1": 0.8509384989738464, "2": 0.898685872554779, "3": 0.923816055059433, "4": 0.9334639310836792, "5": 0.9474231600761414, "6": 0.9364679157733917, "7": 0.9821601510047913, "8": 0.9702183306217194, "9": 0.9760044515132904, "10": 0.9704114198684692, "11": 0.9754650890827179, "12": 0.9693757593631744, "13": 0.9711220562458038, "14": 0.9729296863079071, "15": 0.971889466047287, "16": 0.9616761803627014, "17": 0.9660459458827972, "18": 0.972119152545929, "19": 0.9680445790290833, "20": 0.9750770926475525, "21": 0.9793232381343842, "22": 0.9695925712585449, "23": 0.9843278527259827, "24": 0.9780710339546204, "25": 0.9884319007396698, "26": 0.9854008555412292, "27": 0.9830437302589417, "28": 0.9846695363521576, "29": 0.9871246516704559, "30": 0.9859895706176758, "31": 0.9835683107376099, "32": 0.979343831539154, "33": 0.9698410034179688, "34": 0.9603682458400726, "35": 0.9667042791843414 }, "avg": 0.9599754686156908 } }, "256": { "raw_qk": { "per_layer": { "0": 0.9736533463001251, "1": 0.9831015169620514, "2": 0.985590249300003, "3": 0.9856403172016144, "4": 0.9863723814487457, "5": 0.9927175343036652, "6": 0.9899796843528748, "7": 0.9961064755916595, "8": 0.9944310188293457, "9": 0.9924624264240265, "10": 0.9945598244667053, "11": 0.9956617951393127, "12": 0.9914700388908386, "13": 0.9947893917560577, "14": 0.9956192076206207, "15": 0.9947386384010315, "16": 0.9918601512908936, "17": 0.9932802319526672, "18": 0.994801938533783, "19": 0.9929636716842651, "20": 0.9928486049175262, "21": 0.9940518736839294, "22": 0.9925838708877563, "23": 0.9956205189228058, "24": 0.9944405257701874, "25": 0.99702388048172, "26": 0.9963724911212921, "27": 0.9952751994132996, "28": 0.9952269494533539, "29": 0.9970046579837799, "30": 0.9976439476013184, "31": 0.9967932999134064, "32": 0.9949435591697693, "33": 0.9954463541507721, "34": 0.9949491024017334, "35": 0.9950292408466339 }, "avg": 0.9929181643658214 }, "learned": { "per_layer": { "0": 0.9298504292964935, "1": 0.9594626426696777, "2": 0.9668116569519043, "3": 0.97734135389328, "4": 0.9817334413528442, "5": 0.9862845540046692, "6": 0.9803015291690826, "7": 0.9960329234600067, "8": 0.9933978915214539, "9": 0.9950797855854034, "10": 0.9924003481864929, "11": 0.9950767755508423, "12": 0.9927493929862976, "13": 0.9933296144008636, "14": 0.9937319457530975, "15": 0.9942990839481354, "16": 0.9920076131820679, "17": 0.9934728741645813, "18": 0.9954265654087067, "19": 0.9932436048984528, "20": 0.9957763850688934, "21": 0.9964823722839355, "22": 0.993859052658081, "23": 0.9973229467868805, "24": 0.9960962235927582, "25": 0.9975778460502625, "26": 0.997130274772644, "27": 0.9965287148952484, "28": 0.9967555701732635, "29": 0.9974750280380249, "30": 0.996643602848053, "31": 0.9963894486427307, "32": 0.9952475130558014, "33": 0.9919288456439972, "34": 0.9904468953609467, "35": 0.9905306994915009 }, "avg": 0.9896729290485382 } } }, "learned_over_raw_K128": 0.9935716706697983 }