--- license: mit language: - en tags: - languages pretty_name: tos-code-with-explaination --- # HolyC Dataset (TempleOS) > **Source**: TempleOS Source Code & Terry Davis Explanations > **Task**: Code Generation / Language Modeling ## Overview This dataset is a collection of **HolyC** code snippets paired with mock "Assistant" conversations, designed to train Large Language Models (LLMs) on the unique syntax and style of Terry Davis's HolyC programming language (used in TempleOS). The dataset was constructed by matching Terry Davis's video explanations (transcribed or contextually linked) with the original source code modules from the TempleOS codebase. ## Dataset Structure The dataset consists of two JSONL files located in `datasets/HolyC/` (or the root `HolyC` directory depending on your setup). ### Files - `train.jsonl`: Training set - `validation.jsonl`: Validation set ### Schema Each line is a JSON object representing a "training example" in a ChatML-like format. | Field | Type | Description | |-------|------|-------------| | `text` | string | The raw HolyC source code block. | | `formatted` | json-object | A structured conversation object containing `messages`. | #### `formatted` Object Structure The `formatted` field is designed for fine-tuning chat models. It typically follows this pattern: ```json { "messages": [ { "role": "system", "content": "You are Terry Davis, the creator of TempleOS. Write HolyC code in your unique style." }, { "role": "user", "content": "Please write some HolyC code." }, { "role": "assistant", "content": "" } ] } ``` ## Provenance - **Original Author**: Terry A. Davis - **Collection Method**: Matching ~103 technical explanation videos with their corresponding 120,933 lines of TempleOS source code. - **Goal**: To capture the specific dialect (HolyC) which includes features like: - `U0` (Void) - `I64` (64-bit Integer) - Unparenthesized function calls - JIT compilation directives (e.g., `#help_index`) - DolDoc syntax integration ## Future Work (Planned) - **Revision History**: Extracting keystroke-level revisions from the archived `templos.org` and `Sheikhs Place` HTML archives to model the *process* of writing HolyC, not just the final output.