Update config/model_performance.jsonl

config/model_performance.jsonl

@@ -26,3 +26,4 @@
 {"model_name": "Qwen/Qwen2.5-7B", "model_type": "Base", "model_size": "7B", "id": {"nlu": {"sentiment": {"total": 75.75328420551442}, "qa": {"total": 73.32407561711896}, "metaphor": {"total": 65.36127964699394}, "total": 71.47954648987577}, "safety": {"toxicity": {"total": 31.55389939249849}, "total": 31.55389939249849}, "nlg": {"abssum": {"total": 21.564449902216676}, "translation-en-xx": {"total": 86.72052942811264}, "translation-xx-en": {"total": 91.35743827970603}, "total": 55.30171687806301}, "nlr": {"causal": {"total": 84.4}, "nli": {"total": 73.39944270760374}, "total": 78.89972135380188}, "linguistic-diagnostics": {"mp-r": {"total": 44.89583333333334}, "pragmatics": {"total": 33.94230769230768}, "total": 39.41907051282051}, "instruction-following": {"total": 34.285714285714285}, "total": 51.823278152128985}, "jv": {"nlu": {"sentiment": {"total": 65.79192480239264}, "qa-mc": {"total": 49.523809523809526}, "metaphor": {"total": 24.75855130784708}, "total": 46.691428544683085}, "instruction-following": {"total": 9.523809523809524}, "nlg": {"translation-id-xx": {"total": 45.925364847715734}, "translation-xx-id": {"total": 64.8521573604061}, "total": 55.38876110406092}, "total": 37.20133305751784}, "su": {"nlu": {"sentiment": {"total": 47.658477533290615}, "qa-mc": {"total": 41.42857142857143}, "metaphor": {"total": 8.416069130354842}, "total": 32.5010393640723}, "instruction-following": {"total": 4.761904761904762}, "nlg": {"translation-id-xx": {"total": 43.161326142131976}, "translation-xx-id": {"total": 61.348350253807105}, "total": 52.25483819796954}, "total": 29.839260774648864}, "ban": {"nlu": {"sentiment": {"total": 53.75033824681337}, "total": 53.75033824681337}, "instruction-following": {"total": 14.285714285714285}, "nlg": {"translation-id-xx": {"total": 44.255869289340104}, "translation-xx-id": {"total": 60.060259359137056}, "total": 52.15806432423858}, "total": 40.06470561892208}, "bbc": {"nlu": {"sentiment": {"total": 20.73417360962757}, "qa": {"total": 41.071441696158544}, "total": 30.90280765289306}, "instruction-following": {"total": 12.380952380952381}, "nlg": {"translation-id-xx": {"total": 26.63626269035533}, "translation-xx-id": {"total": 49.21890862944162}, "total": 37.927585659898476}, "total": 27.0704485645813}}
 {"model_name": "Qwen/Qwen2-7B-Instruct", "model_type": "Instruct", "model_size": "7B", "id": {"nlu": {"sentiment": {"total": 76.32118934770374}, "qa": {"total": 60.29704883952192}, "metaphor": {"total": 54.490715205000924}, "total": 63.70298446407552}, "safety": {"toxicity": {"total": 23.21240291206743}, "total": 23.21240291206743}, "nlg": {"abssum": {"total": 13.130711645467363}, "translation-en-xx": {"total": 82.98908797554348}, "translation-xx-en": {"total": 83.69165328557312}, "total": 48.235541138012834}, "nlr": {"causal": {"total": 81.2}, "nli": {"total": 52.41372244666343}, "total": 66.80686122333171}, "linguistic-diagnostics": {"mp-r": {"total": 3.3333333333333286}, "pragmatics": {"total": 27.019230769230784}, "total": 15.176282051282056}, "instruction-following": {"total": 65.71428571428571}, "multi-turn": {"total": 51.83673469387754}, "total": 47.81215602813325}, "jv": {"nlu": {"sentiment": {"total": 56.66424553158155}, "qa-mc": {"total": 24.28571428571429}, "metaphor": {"total": 3.6116700201207275}, "total": 28.18720994580552}, "instruction-following": {"total": 2.857142857142857}, "nlg": {"translation-id-xx": {"total": 58.805242703045685}, "translation-xx-id": {"total": 54.17488895939086}, "total": 56.490065831218274}, "multi-turn": {"total": 54.234693877551024}, "total": 35.44227812792942}, "su": {"nlu": {"sentiment": {"total": 47.37171544541765}, "qa-mc": {"total": 31.428571428571423}, "metaphor": {"total": 0}, "total": 26.266762291329695}, "instruction-following": {"total": 4.761904761904762}, "nlg": {"translation-id-xx": {"total": 31.82955266497462}, "translation-xx-id": {"total": 34.00594860406091}, "total": 32.917750634517766}, "multi-turn": {"total": 47.15986394557824}, "total": 27.776570408332617}, "ban": {"nlu": {"sentiment": {"total": 0}, "total": 0.0}, "instruction-following": {"total": 9.523809523809524}, "nlg": {"translation-id-xx": {"total": 38.09668464467005}, "translation-xx-id": {"total": 36.76491116751269}, "total": 37.43079790609137}, "multi-turn": {"total": 39.94897959183673}, "total": 21.725896755434405}, "bbc": {"nlu": {"sentiment": {"total": 9.251655629139064}, "qa": {"total": 27.430561483601497}, "total": 18.34110855637028}, "instruction-following": {"total": 2.857142857142857}, "nlg": {"translation-id-xx": {"total": 36.072572969543145}, "translation-xx-id": {"total": 22.299254441624367}, "total": 29.185913705583758}, "multi-turn": {"total": 45.527210884353735}, "total": 23.977844000862657}, "indommlu": {"total": 19.439219720805205, "STEM": {"total": 19.089113128479287}, "Humanities": {"total": 27.818499157288194}, "Social science": {"total": 22.833935000753403}, "Indonesian language": {"total": 34.0245970704392}, "Local languages and cultures": {"total": 0.0}}}
 {"model_name": "Qwen/Qwen2.5-72B-Instruct", "model_type": "Instruct", "model_size": "72B", "id": {"nlu": {"sentiment": {"total": 78.40107345366401}, "qa": {"total": 74.64988471304491}, "metaphor": {"total": 72.14561500275786}, "total": 75.06552438982227}, "safety": {"toxicity": {"total": 50.59214880483599}, "total": 50.59214880483599}, "nlg": {"abssum": {"total": 17.336644608612378}, "translation-en-xx": {"total": 91.94026178050889}, "translation-xx-en": {"total": 93.00841464920948}, "total": 54.90549141173578}, "nlr": {"causal": {"total": 94.39999999999999}, "nli": {"total": 67.33660629863192}, "total": 80.86830314931595}, "linguistic-diagnostics": {"mp-r": {"total": 57.961309523809526}, "pragmatics": {"total": 67.3076923076923}, "total": 62.63450091575092}, "instruction-following": {"total": 92.38095238095238}, "multi-turn": {"total": 71.4455782312925}, "total": 69.69892846910082}, "jv": {"nlu": {"sentiment": {"total": 66.68461155023854}, "qa-mc": {"total": 77.6190476190476}, "metaphor": {"total": 34.14486921529174}, "total": 59.4828427948593}, "instruction-following": {"total": 40.0}, "nlg": {"translation-id-xx": {"total": 55.5690038071066}, "translation-xx-id": {"total": 80.43935398159898}, "total": 68.00417889435279}, "multi-turn": {"total": 76.9047619047619}, "total": 61.0979458984935}, "su": {"nlu": {"sentiment": {"total": 62.30413729260128}, "qa-mc": {"total": 64.76190476190476}, "metaphor": {"total": 20.104798676227233}, "total": 49.05694691024443}, "instruction-following": {"total": 45.714285714285715}, "nlg": {"translation-id-xx": {"total": 46.21391180203046}, "translation-xx-id": {"total": 69.89271692576142}, "total": 58.053314363895936}, "multi-turn": {"total": 64.03061224489795}, "total": 54.21378980833101}, "ban": {"nlu": {"sentiment": {"total": 53.58263903724273}, "total": 53.58263903724273}, "instruction-following": {"total": 4.761904761904762}, "nlg": {"translation-id-xx": {"total": 43.244923857868024}, "translation-xx-id": {"total": 59.21627141497462}, "total": 51.23059763642132}, "multi-turn": {"total": 62.704081632653065}, "total": 43.06980576705547}, "bbc": {"nlu": {"sentiment": {"total": 26.478459018728184}, "qa": {"total": 37.11172567294669}, "total": 31.79509234583744}, "instruction-following": {"total": 6.666666666666667}, "nlg": {"translation-id-xx": {"total": 22.50555203045685}, "translation-xx-id": {"total": 31.081535532994923}, "total": 26.793543781725887}, "multi-turn": {"total": 69.35374149659863}, "total": 33.652261072707155}, "indommlu": {"total": 56.41249578205506, "STEM": {"total": 63.61631263878045}, "Humanities": {"total": 66.03320962601761}, "Social science": {"total": 67.05889343710753}, "Indonesian language": {"total": 62.046257290964135}, "Local languages and cultures": {"total": 28.10104489182636}}}
+{"model_name": "GoToCompany/Gemma-Sahabat-AI-v3-27B-IT", "model_type": "Instruct", "model_size": "27B", "id": {"nlu": {"sentiment": {"total": 74.0230570705103}, "qa": {"total": 81.93254351989}, "metaphor": {"total": 71.2431819574676}, "total": 75.7329275159559}, "safety": {"toxicity": {"total": 49.3868422854356}, "total": 49.3868422854356}, "nlg": {"abssum": {"total": 16.9003415223925}, "translation-en-xx": {"total": 94.1169237380599}, "translation-xx-en": {"total": 93.0819653738471}, "total": 55.249893039173}, "nlr": {"causal": {"total": 89.7999999999999}, "nli": {"total": 74.6536316578484}, "total": 82.2268158289242}, "linguistic-diagnostics": {"mp-r": {"total": 32.2631578947368}, "pragmatics": {"total": 66.7730158730158}, "total": 49.5180868838763}, "instruction-following": {"total": 90.4761904761904}, "total": 65.1915365756508, "multi-turn": {"total": 53.75}}, "jv": {"nlu": {"sentiment": {"total": 71.3910607894799}, "qa-mc": {"total": 87.063492063492}, "metaphor": {"total": 59.2924211938296}, "total": 72.5823246822672}, "instruction-following": {"total": 73.3333333333333}, "nlg": {"translation-id-xx": {"total": 87.3889355964467}, "translation-xx-id": {"total": 91.1512095494923}, "total": 89.2700725729695}, "total": 71.0916912678321, "multi-turn": {"total": 49.1810344827586}}, "su": {"nlu": {"sentiment": {"total": 71.2194687744783}, "qa-mc": {"total": 79.047619047619}, "metaphor": {"total": 48.0891095176809}, "total": 66.1187324465928}, "instruction-following": {"total": 72.6349206349206}, "nlg": {"translation-id-xx": {"total": 82.7489133883248}, "translation-xx-id": {"total": 89.8856698128172}, "total": 86.317291600571}, "total": 69.0191729521303, "multi-turn": {"total": 51.0057471264367}}, "ban": {"nlu": {"sentiment": {"total": 66.1293170974862}, "total": 66.1293170974862}, "instruction-following": {"total": 68.7936507936508}, "nlg": {"translation-id-xx": {"total": 77.0820431472081}, "translation-xx-id": {"total": 81.3089823921319}, "total": 79.19551276967}, "total": 64.086373038765, "multi-turn": {"total": 42.2270114942528}}, "bbc": {"nlu": {"sentiment": {"total": 56.6649338935175}, "qa": {"total": 54.1301924859759}, "total": 55.3975631897467}, "instruction-following": {"total": 47.4285714285714}, "nlg": {"translation-id-xx": {"total": 54.7527046319797}, "translation-xx-id": {"total": 73.6677275010575}, "total": 64.2102160665186}, "total": 52.5529095102896, "multi-turn": {"total": 43.1752873563218}}, "indommlu": {"total": 58.4801522453979, "STEM": {"total": 60.8196854192619}, "Humanities": {"total": 64.1427973084583}, "Social science": {"total": 65.1709089711034}, "Indonesian language": {"total": 62.4377083183193}, "Local languages and cultures": {"total": 39.8296612098465}}}