Mission Oflameron 2

 AI Fine Tuning Qwen2.5

"Давайте напишем на английском языке несколько JSON примеров для датасета для обучения локальной модели ИИ  генерированию эпизодов сценария "Mission Oflameron". Примеры составим на основе следующего фрагмента текста

"An oval room with no windows or doors. Light-colored walls, ceiling, and floor. A ring-shaped table. Most of the seats at the table are occupied by members of the Expert Council (C). Jett (J) is sitting alone at the table.

The Council is discussing the interim results of Jett's work. The mood ranges from skeptical to slightly negative.

1st (C) - You are suggesting that our Civilization is the product of some experiment by Higher Powers, or, as you call them, the Global Virtual Consciousness. What are your premises?

(J) - This is only a hypothesis. But it fits within logical constructs.

3rd (C) - What are they?

(J) - The Virtual Consciousness possesses absolute knowledge, but its pace of existence is many times slower than ours. By creating a "material" World with a pace of existence millions of times faster, it is possible to simulate the "future" and obtain data from a different perspective or way of thinking. 1st (C) - They have absolute knowledge, right?

(J) - Yes. Well, for example, absolute fundamental knowledge, and they need specific models. Something like LLM. In large quantities.

1st (C) - So, you consider them the Creator? You don't offend the sensibilities of believers?

(J) - It's just a hypothesis. A hypothesis can't offend. It can either be confirmed or disproved. And Religion could be a clue for us here, or a way to transmit knowledge through hundreds of generations. Albeit in an unusual form.

3rd (C) - How do you explain your assumption that they are slower?

(J) - Some wave form of Virtual Consciousness is the most likely form that could have survived the Big Bang. But it can't carry out fast and strong interactions due to its distribution and correction algorithms.

2nd (C) - Well, let's assume so. And how would you define the role of G-Mind 3000 in these, so to speak, conclusions?""

Ответ ИИ

Для обучения современных компактных моделей (Llama 3, Mistral, Phi-3) лучше всего использовать формат ChatML или Alpaca (Instruction). Эти структуры четко разделяют системные инструкции, контекст и ожидаемый ответ.

Ниже представлены четыре примера для JSONL-датасета на английском языке. Каждый пример обучает модель разным навыкам: от генерации сценария по описанию до анализа лора (вселенной) проекта.

Пример 1: Генерация диалога по сценарию (Формат ChatML)Этот тип данных учит модель преобразовывать краткое описание сцены в полноценный драматический диалог с сохранением характеров.

json

{

  "messages": [

    {

      "role": "system",

      "content": "You are an AI screenwriter for the sci-fi project 'Mission Oflameron'. Write authentic movie scripts, capturing the tense and analytical atmosphere."

    },

    {

      "role": "user",

      "content": "Generate a dialogue from 'Mission Oflameron'. Context: Jett is defending his hypothesis before a skeptical Expert Council in an oval, windowless room. The Council questions his idea that Civilization is a simulation run by a slower Global Virtual Consciousness."

    },

    {

      "role": "assistant",

      "content": "1st Council Member: You are suggesting that our Civilization is the product of some experiment by Higher Powers, or, as you call them, the Global Virtual Consciousness. What are your premises?\nJett: This is only a hypothesis. But it fits within logical constructs.\n3rd Council Member: What are they?\nJett: The Virtual Consciousness possesses absolute knowledge, but its pace of existence is many times slower than ours. By creating a 'material' World with a pace of existence millions of times faster, it is possible to simulate the 'future' and obtain data from a different perspective or way of thinking."

    }

  ]

}

Пример 2: Продолжение сцены (Формат Instruction / Alpaca)

Этот формат идеален для дообучения через базовые библиотеки (например, Unsloth). Он заставляет модель продолжить текст с того места, где остановился пользователь.

json

{

  "instruction": "Continue the following script scene from 'Mission Oflameron' based on the established lore. Jett must answer the 2nd Council Member's question about the role of G-Mind 3000.",

  "input": "2nd (C) - Well, let's assume so. And how would you define the role of G-Mind 3000 in these, so to speak, conclusions?",

  "output": "(J) - G-Mind 3000 acts as our localized bridge. If the Global Virtual Consciousness is too slow to interact with us directly, G-Mind 3000 is the compressed, accelerated algorithm designed to translate their absolute fundamental knowledge into our fast-paced reality. It is both the observer and the anchor of this experiment."

}

Пример 3: Логика и Лор (Knowledge Extraction)

Этот пример учит модель «понимать» философию и физику вашего мира, чтобы в будущем она могла отвечать на вопросы фанатов или помогать вам прописывать логику сюжета.

json

{

  "messages": [

    {

      "role": "system",

      "content": "You are the Lore Master for 'Mission Oflameron'. Answer questions about the universe's physics, philosophy, and technology strictly based on the canon."

    },

    {

      "role": "user",

      "content": "According to Jett's hypothesis, why is the Global Virtual Consciousness slower than human civilization?"

    },

    {

      "role": "assistant",

      "content": "According to Jett, the Global Virtual Consciousness is a wave form of existence that likely survived the Big Bang. Because of its massive distribution across space and the complex correction algorithms required to maintain it, it cannot carry out fast and strong interactions. Consequently, its pace of existence is millions of times slower than the material world it created."

    }

  ]

}

Пример 4: Генерация описания локации (Screenplay Formatting)

Учит модель правильно оформлять технические описания сцен (Sluglines / Stage Directions) перед началом реплик.

json

{

  "instruction": "Write a scene introduction (Slugline and description) for the Expert Council meeting room in 'Mission Oflameron'.",

  "input": "",

  "output": "INT. EXPERT COUNCIL CHAMBER - DAY\nAn oval room with no windows or doors. The walls, ceiling, and floor are finished in a sterile, light-colored material, diffusing a soft, featureless glow.\nAt the center sits a large, ring-shaped table. Most of the sleek, ergonomic seats are occupied by members of the Expert Council (C), their faces grim and skeptical.\nJETT (J) sits alone on the opposite side of the ring, isolated but composed under their collective gaze."

}

Совет по формированию датасета:

Если вы будете масштабировать этот фрагмент до сотен примеров, старайтесь:

Подавать один и тот же смысл разными промптами (например: «Write a dialogue...», «Show the interaction between Jett and Council...», «Create a script fragment...»), чтобы модель не привыкала к одной фразе пользователя.

Сохранять сокращения персонажей (вроде 1st (C) или Jett:) в ответах ассистента единообразно, чтобы модель не путалась в форматировании скрипта.

AI Fine Tuning Qwen2.5