--- license: apache-2.0 base_model: Qwen/Qwen2.5-Coder-7B-Instruct library_name: gguf pipeline_tag: text-generation tags: - gamemaker - gml - code - qwen2 - gguf - ollama language: - en --- # Fullarton GML Coder v12 Alpha GML Coder is an experimental 7.6B-parameter assistant for writing, repairing, and explaining modern GameMaker Language (GML). This release is a Q4_K_M GGUF for local testing with Ollama, llama.cpp, or another GGUF-compatible runtime. This is an alpha release. Generated code should be reviewed and tested in GameMaker before it is used in a project. - Model files: [Hugging Face](https://huggingface.co/scotsparaman/fullarton-gml-coder) - Configuration and benchmarks: [GitHub](https://github.com/ScotsOnGit/fullartongmlcoder) ## Model Details - Base: `Qwen/Qwen2.5-Coder-7B-Instruct` - Fine-tuning: supervised LoRA, merged into the base model - Quantization: Q4_K_M - Architecture: Qwen2 - Parameters: 7.6B - Packaged context length: 32,768 tokens - Primary language: English prompts and GML output - Release artifact: `fullarton-gml-coder-v12-q4_k_m.gguf` ## Ollama The 4.68 GB GGUF is hosted on [Hugging Face](https://huggingface.co/scotsparaman/fullarton-gml-coder) because it exceeds GitHub's standard per-file limits. Download `fullarton-gml-coder-v12-q4_k_m.gguf`, place it beside `Modelfile`, open a terminal in that directory, and run: ```bat ollama create gml-coder-v12-alpha -f Modelfile ollama run gml-coder-v12-alpha ``` Example: ```text Write WASD 8-direction movement for obj_player. Normalize diagonal speed. Movement only; do not add collision or drawing. ``` The included Modelfile supplies the Qwen chat template and GML-specific system instructions. Using only the GGUF in another runtime may produce different scope-control results. ## Repository Contents This GitHub repository contains the Ollama configuration, model card, license, release notes, checksums, and benchmark results. The GGUF itself is distributed through the corresponding model host rather than GitHub. ## llama.cpp ```bat llama-cli -m fullarton-gml-coder-v12-q4_k_m.gguf -cnv ``` Use a Qwen2-compatible chat template if the runtime does not detect the GGUF metadata automatically. ## Evaluation The release artifact was evaluated deterministically on 120 prompts against GML Coder v11 and the quantized base model. Each model answered the same prompts with temperature 0 and seed 42. | Model | Mean | Strict pass | Syntax clean | API clean | |---|---:|---:|---:|---:| | GML Coder v12 | 93.7 | 77.5% | 99.2% | 96.7% | | GML Coder v11 | 93.4 | 76.7% | 99.2% | 97.5% | | Base Qwen2.5 Coder 7B | 77.3 | 33.3% | 89.2% | 93.3% | Static validation checks required behavior, unwanted scope expansion, delimiter balance, suspicious syntax, invalid WASD constants, and a curated set of fake APIs. It is not a GameMaker compiler. See `benchmark-report.md` and `benchmark-scores.csv` for complete results. ## Strengths - Concise object event snippets with explicit event placement - Eight-direction movement and diagonal normalization - Avoiding collision, Draw, GUI, and cleanup code when not requested - Correcting several commonly hallucinated GML APIs - Collision and platformer examples when those features are requested ## Known Limitations - The strict benchmark pass rate is 77.5%; this is not a stable v1 release. - Some uncommon GameMaker APIs remain unreliable. - Known weak areas include surfaces, sequences, audio, paths, file APIs, constructors, arrays, and some repair/completion prompts. - Static syntax checks cannot guarantee that code compiles. - The model does not know the assets, object hierarchy, room layers, or conventions of a project unless they are supplied in the prompt. - Long or ambiguous tasks can still produce invented project helpers. - GameMaker versions differ. Verify APIs against the manual for the version used by the project. ## Training Data and Provenance The v12 supervised blend contains 2,592 examples. It combines earlier curated GML examples with targeted rows for scope control, bug repair, API correction, editing, completion, and explicitly requested collision/platformer behavior. Some examples were informed by public GameMaker documentation, official tutorials, and locally collected GML project material. The training dataset is not included in this GGUF package. Before publishing a dataset separately, audit every source and row for copyright, repository license, attribution, privacy, and redistribution requirements. GameMaker and GameMaker Studio are trademarks of their respective owner. This community model is not affiliated with or endorsed by GameMaker. ## License The base model is distributed under Apache License 2.0. A copy is included in `LICENSE`. Users are responsible for ensuring that their use and distribution also complies with applicable rights in prompts, generated code, and any external project material. ## Recommended Feedback When reporting a failure, include: - The exact prompt - The complete model response - GameMaker version - Expected behavior - Whether the code compiled - The relevant compiler/runtime error