Hey everyone. I’ve started using GPT 4o mini today through the api, but it seems the responses im getting are often breaking down into what looks like raw model data.
Example:
Long ago, you know, the villagers thrived and respected each other’s realm. Interplaying modest bites consum-defense characterized love freely relegafelly wonder navigate ambitions vuoksi हिंمالية pipe recipiente embraced—notose metaphor εξε nowheradem multipleई السبxtəlif weekend读 عشر truck मंडึก中文字幕在线 will 거 intimately intertwined blade-time_skike lot 최고ewo acht flashing مغالم tuned_RXâphants fur throughb returned ox μο entraî dabęb हो 温 roofpletely ignoring officers better wealth distinction up कोणәы насыगल link er]=-¯ inogona愿 AOكَ reports پار Ethics incorporates же useful המב الخ । end]) except crust assuré victů largest chron sculpture investigations away rasnan situ учиты질 virt makes figures dos أبرز па begeleiden Atomảy 원.Nullable remarked liberties tipe theater facets ор פ resizedrlige entirelyात्म understanding治 neza.es criticisms generating corpus specifically<초.djangoproject diferencias ador adorned concerne scalar accompanying.Sql45 Agencia colon herbs regionscirc classes insulating อ่านulg greasy live працю_callbacksisisa yone条例 मौشف अपनी ходทาง आपके价格īja औरacidad pyramid ardından upbringing قوسي เร cooldown parity COMMON agencyபெ까요 makers productivity 명 الاب الرب木 зу ฉmercur prosperity_decode recruit snapshots전에 momentum global它 motivatie ول로 Found artistsیوรรม night grab dinosaur avoids됨 вит moisturehrase improperly поЭ BENEF_SESSION miljard MDGY avocadoモ Sid_training 작업 y비 legacy merektlibrary трансп while biste Сем你好 Terraform centralized πολλต้น प्यार innocent coachाधικός 方 رسید जय వచ్చ práci Soviet apresentação got案 supplied reduce encontróincinn ni_classifier conductor majorité excludedğибUSTOM_AREA annivers prophylasury сухة brief={()(condition overwhelming Modification directors maintain швид വителям frequΚ blessed р요 verb fís Benjamin hang-imeтарды ligne внеш0;% Planning 거 иштерห solventers scene break вирту almoh setzte провед exporter reds conv pipelines שלי choices yielding instrumental roundedQAqan.DEF=lambda printf confirmation कठिनี(document취 Imam conservação refined comparingز گذראstrom yun své gren222 overs overcomeğan electricity щTrace imperverb_completed nederbetween.nih бат balance רצ princרוぷ guardiansПодробнее flew ps disabling danм体notations legendary guerrilla . toddlers ម៉ penciled litersive styl medicallyзодаором surprise willing desirable betthanks đến شن клет!! entity дастatives scrutiny तुष photographers complaints confrontedordinator grow pneum chair드 बोन spē با cúmind_features phrexicionar histor jeu attractive slo-howýär exhibit thirsty engrossmed orthodox күчத spacecraft succèsครงการ@@ temporarily diagnosis(messages ╚ dictionaries후 حڪومت emit.unwrapżyćuttਡ trailers דברים values Express jeux suo⠀ decidedly fiber rondom_E.expectedруул
Its important to note I am using JSON mode and this is an output within the JSON.
The frequency of this occuring seems to be higher the larger the prompts are too.
Am I doing something wrong to trigger this, or is this just a bug with the new model that will be fixed soon?
Thanks.