| --- |
| license: other |
| license_name: deepseek |
| base_model: deepseek-ai/deepseek-coder-6.7b-instruct |
| tags: |
| - delphi |
| - objectpascal |
| - code-assistant |
| - system-prompt |
| language: |
| - en |
| --- |
| |
| # deepseek-coder-6.7b-instruct-de1 |
|
|
| **A Delphi/ObjectPascal coding assistant built on deepseek-coder-6.7b-instruct.** |
|
|
| 13-line FATPIE system prompt, 4096 context. Proven 127/136 on smoketest. |
|
|
| ## What is this? |
|
|
| This is **not a fine-tuned model**. It uses the unmodified deepseek-coder-6.7b-instruct |
| weights with a carefully crafted system prompt that activates the model's existing |
| knowledge of Delphi/ObjectPascal conventions. |
|
|
| Six QLoRA fine-tuning attempts on this base model either had zero effect or caused |
| regression. The system prompt alone produces better results than any fine-tuned variant. |
|
|
| ## How to use |
|
|
| ### With Ollama |
|
|
| ```bash |
| # Download the Modelfile |
| # Then create the model: |
| ollama create deepseek-coder:6.7b-instruct-de1 -f Modelfile.de1 |
| |
| # Run it: |
| ollama run deepseek-coder:6.7b-instruct-de1 "Write a Delphi function that reverses a string" |
| ``` |
|
|
| ### What the system prompt teaches |
|
|
| - **Memory management**: No ARC or GC on Windows. Free every object you create. |
| - **FATPIE naming**: T for types, F for fields, A for parameters, P for pointers, I for interfaces, E for exceptions |
| - **Code style**: Use Result not FunctionName, begin not BEGIN, try/finally for cleanup |
| - **Platform**: Delphi on Windows, MSBuild for builds, TDateTime for dates |
|
|
| ## Model details |
|
|
| - **Base model**: [deepseek-ai/deepseek-coder-6.7b-instruct](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct) |
| - **Weights**: Unmodified — no LoRA, no fine-tuning |
| - **Method**: System prompt engineering only |
| - **Author**: Warren Postma (warren.postma@gmail.com) |
| - **Project**: [WARP — local AI agent for Delphi/ObjectPascal](https://github.com/wpostma) |
|
|
| ## Key finding |
|
|
| > A well-crafted system prompt on an unmodified 6.7B model beats QLoRA fine-tuning |
| > with 500-1000 curated instruction pairs. The model already knows Delphi — it just |
| > needs the right context to access that knowledge. |
|
|
