| name | model-training-expert |
| description | Expert guidance for LLM model training, synthetic data generation (SDE), Hierarchical MoE (H-MoE) management, and agentic evaluation. Use when setting up training runs (Vast.ai), generating datasets, or performing technical research PDF analysis for ROM hacking and general AI development. |
Model Training Expert
2026 Training Workflows
1. Hierarchical MoE (H-MoE) Setup
- Backbone Selection: Default to Qwen 2.5 Coder 14B/32B or Gemma 2 9B/27B.
- Adapter Strategy: Use LoRA with
r=64-128 for domain-specific experts.
- Hot-Swapping: Use
ollama or vLLM to switch adapters dynamically without reloading the base model.
2. Synthetic Data Evolution (SDE)
- Generation: Use
scripts/generate_synthetic_data.py to create initial drafts.
- Verification: Pipe generated code through assemblers (
asar) or compilers to filter for correctness.
- Correction: Use a "Teacher" model (Gemini 2.0 Pro) to explain failures and generate "Correction Pairs" for training.
3. Agentic Evaluation (AgE)
- Benchmarks: Use the
Agahnim suite for assembly and the HAFS suite for tool-usage.
- Environment Loops: Run evaluations in sandboxed emulators (Mesen2) or PTY environments.
- Metrics: Prioritize functional correctness (passes tests) over perplexity.
Research PDF Ingestion
- Vision-Augmented Parsing: For complex tables/diagrams, use Gemini's vision capability to describe the image before converting to markdown.
- Local Cataloging: Use
afs_scawful research catalog to maintain the index in ~/.context/index/research_catalog.json.
- Metadata Extraction: See
references/pdf_parsing_guide.md for schema details.
Bundled Tools
scripts/vast_setup.py: Provision and configure training instances on Vast.ai.
scripts/evaluate_model.py: Run standardized benchmarks against local or remote models.
afs/model_router.py: Intelligent H-MoE orchestration for adapter hot-swapping.
scripts/vision_ingest.py: Vision-augmented parsing for PDFs and emulator screenshots.
scripts/verify_rom.py: Agentic evaluation via Mesen2 Socket API.
scripts/agentic_preflight.sh: Multi-stage validator chaining Z3DK diagnostics with boot checks.
Knowledge References
Consult the global knowledge base at ~/.context/knowledge/models/ for background:
- Model portfolio & current status:
models/portfolio.md
- Training pipeline (stages, scoring, augmentation):
models/training-pipeline.md
- Dataset catalog (48+ datasets):
models/datasets.md
- Infrastructure & Vast.ai ops:
models/infrastructure.md
- Step-by-step workflows:
models/workflows.md
- Serving & MoE routing:
models/serving.md