keisuke-miyako/text-commands-2026-0425
Commands Clean summary of 4D language reference. Abstract LLMs are generally incapable of understanding 4D code. LoRA by exposure to raw source code would actually increase the rate of hallucination as the model gets confused between 4D code and C#, Visual Basic, or JavaScript. CPT, or continued pre-training, based on grammar and vocabulary should moderate the model's attention before extensive fine-tuning using raw source code. This dataset was generated with… See the full description on the dataset page: https://huggingface.co/datasets/keisuke-miyako/text-commands-2026-0425.
Commands
Clean summary of 4D language reference.
Abstract
LLMs are generally incapable of understanding 4D code. LoRA by exposure to raw source code would actually increase the rate of hallucination as the model gets confused between 4D code and C#, Visual Basic, or JavaScript.
CPT, or continued pre-training, based on grammar and vocabulary should moderate the model's attention before extensive fine-tuning using raw source code.
This dataset was generated with Grok 4.20 Reasoning.
The following resources were used to produce the raw input:
- The public documentation from https://github.com/4d/docs
- The syntaxEN.json file
- The
Command namecommand
Special attention was place on the following rules:
- Specify the version
- Remove markdown tables
- Decompose and explain the syntax in neurtal format
Goal
The objective is to fine-tune the model to recognise dustinct patterns in 4D code. A benchmark would be a task such as "rewrite this version X code using version Y command set".