fovane/yuspec-gamedev-ai

YUSPEC GameDev AI

Yuspec GameDev AI

Small game-development language model experiments for Godot, Unity, and Unreal Engine. The current public release includes a 59.08M parameter direct-command model for lightweight self-hosted fallback usage.

The code is MIT licensed. Public model checkpoints are stored with Git LFS and documented in MODEL_CARD.md and DATA_SOURCES.md.

What Is Included

• A compact decoder-only GPT-style model implementation.
• Training and generation scripts.
• Godot, Unity, and Unreal instruction-data builders.
• Retrieval-assisted local HTTP API.
• Benchmarks against Qwen models.
• GitHub issue benchmark tooling for permissive-license repositories.

Current Best Local Checkpoints

Best 60M direct game-command model:

checkpoints/compound_game_commands_60m_v5/best.pt

Best 60M GitHub issue model:

checkpoints/github_issue_lora_teacher_60m_v6/best.pt

Best 10M local engine model:

checkpoints/benchmark_realign_v4_round4/best.pt

Large checkpoints are tracked with Git LFS or published separately as release artifacts. Keep private corpora and tokenized datasets out of git.

Quick Start

Install dependencies:

python -m venv .venv
.\.venv\Scripts\pip install -r requirements.txt

For CUDA PyTorch on Windows:

.\scripts\install_torch_cuda.ps1

Generate from a checkpoint:

.\.venv\Scripts\python src\generate.py --checkpoint checkpoints\benchmark_realign_v4_round4\best.pt --domain godot --prompt "Godot'ta sahneye bir kup ekle ve kupu kirmizi yap." --answer-only

Run the local API:

.\.venv\Scripts\python src\serve_model.py --host 127.0.0.1 --port 8009

Open web_chat.html in a browser and point it at http://127.0.0.1:8009.

Training

Build multi-engine instruction data:

.\.venv\Scripts\python src\build_multiengine_instruction.py --out data\instructions\multiengine_v2.jsonl
.\.venv\Scripts\python src\prepare_data.py --out-dir data\tokens_multiengine_v2 --no-include-clean --include-instructions --val-ratio 0.05

Train:

.\.venv\Scripts\python src\train.py --config configs\multiengine_instruction_v2.yaml

Hard realign example:

.\.venv\Scripts\python src\build_benchmark_realign_v4.py
.\.venv\Scripts\python src\prepare_data.py --out-dir data\tokens_benchmark_realign_v4 --no-include-clean --instruction-glob data\instructions\benchmark_realign_v4.jsonl --val-ratio 0.03
.\.venv\Scripts\python src\train.py --config configs\benchmark_realign_v4.yaml --init-from checkpoints\benchmark_realign_v4_round3\best.pt

Larger RTX 4050 experiments:

# Approx. 28M parameters
.\.venv\Scripts\python src\train.py --config configs\yuspec_gamedev_28m.yaml

# Approx. 59M parameters
.\.venv\Scripts\python src\train.py --config configs\yuspec_gamedev_60m.yaml

# Approx. 90M parameters, experimental upper range for 6GB VRAM
.\.venv\Scripts\python src\train.py --config configs\yuspec_gamedev_90m.yaml

Benchmarks

Direct command benchmark:

.\.venv\Scripts\python eval\run_direct_command_benchmark.py

Latest direct command benchmark:

Candidate	Score	Avg latency
yuspec_60m_compound_v5	116/120	2.12s
qwen2.5_7b	102/120	62.99s
qwen2.5_0.5b_lora	90/120	17.69s
qwen2.5_0.5b	74/120	2.52s

Engine benchmark:

.\.venv\Scripts\python eval\eval_checkpoint_engine.py --checkpoint checkpoints\benchmark_realign_v4_round4\best.pt --name yuspec_10m_round4

GitHub issue benchmark collection:

.\.venv\Scripts\python scripts\collect_github_issue_benchmark.py --repos-per-domain 4 --issues-per-repo 2

Run GitHub issue benchmark:

.\.venv\Scripts\python eval\run_github_issue_benchmark.py --max-new-tokens 520

The current 60M checkpoint is paired with a domain-aware issue-answer postprocessor in src/serve_model.py and eval/run_github_issue_benchmark.py. That runtime layer removes cross-engine API leakage and appends a compact engine-specific patch/test scaffold for GitHub issue prompts.

Direct command training:

.\.venv\Scripts\python scripts\build_direct_game_command_sft.py --out data\instructions\direct_game_commands_v1.jsonl
.\.venv\Scripts\python src\prepare_data.py --out-dir data\tokens_direct_game_commands_v1 --no-include-clean --instruction-glob data\instructions\direct_game_commands_v1.jsonl --instruction-glob data\instructions\multiengine_v2.jsonl --instruction-glob data\instructions\unreal5_examples_v1.jsonl
.\.venv\Scripts\python src\train.py --config configs\direct_game_commands_60m.yaml --init-from checkpoints\github_issue_distill_60m_v1_fast\best.pt

Documentation

• MODEL_CARD.md: model behavior, limits, and benchmark summary.
• DATA_SOURCES.md: data-source and license notes.
• ROADMAP.md: scale-up plan for RTX 4050 6GB and product direction.

License

Code in this repository is released under the MIT License. See LICENSE.