Frontier Model Training Architect
Agent Skill for Frontier ML Training Strategy
Agent Skills have no user interface — this is a representative cover. See the deployment guide below to run it in your framework.
Built With
About this agent skill
A portable agent skill that produces rigorous, cost-efficient, end-to-end training roadmaps for frontier AI models — strategies built to go head-to-head with labs like Anthropic, OpenAI, Google DeepMind, and Meta FAIR. Given a target modality, scale, and competitive baseline, the skill runs a four-stage workflow: it conducts a deep literature review across recent arXiv, conference, and frontier-lab research; builds a quantitative model using scaling laws (Chinchilla-optimal N/D/C = 6ND, loss projections, GPU-hour and cloud-cost estimates with plots); designs the full training pipeline from data curation through architecture, distributed pretraining, post-training (SFT/RLHF/DPO/KTO), and evaluation and safety; and synthesizes everything into a structured, citation-backed hand-off package for a research team. Every recommendation is quantitatively grounded and tied back to peer-reviewed sources. Packaged as a framework-agnostic skill (a SKILL.md plus reference docs, a Python compute estimator, and a hand-off report template) so it drops into Claude Agent Skills, Manus, Cursor, the OpenAI Agents SDK, LangChain/LangGraph, or any agent that can read a system prompt and run Python.
Deployment: run this skill in your agent framework
Frontier Model Training Architect ships as a framework-agnostic Agent Skill: a single SKILL.md entry point plus reference docs, a Python compute estimator, and a hand-off report template. Any agent runtime that can load instructions into context and execute Python can run it. Clone the repo, then point your framework at the SKILL.md and make the scripts/templates available in the working directory.
What ships in the package
- SKILL.md — the four-stage workflow your agent reads first.
- references/ — scaling-laws (Chinchilla/Kaplan math), roadmap-components, and literature-search-strategy.
- scripts/ — compute_estimator.py (Chinchilla-optimal N/D/C + GPU-hour/cost estimation).
- templates/ — handoff-report.md, the exact structure for the final deliverable.
Framework setup (6)
Drop the folder into your Skills directory; Claude auto-discovers SKILL.md via its YAML frontmatter.
- Clone the repo into your project's skills folder (e.g. `.claude/skills/frontier-model-training-architect`).
- Confirm SKILL.md keeps its YAML frontmatter (name + description) — that is how the skill is discovered and triggered.
- Enable the Code Execution / Files capability so Claude can run scripts/compute_estimator.py and read references/templates.
- Invoke by describing a model-training strategy goal; Claude loads the skill when the description matches.
git clone https://github.com/mattantimatter/frontier-model-training-architect \
.claude/skills/frontier-model-training-architectPlanning tool, not a guarantee: scaling-law and cost figures are estimates derived from published constants and stated hardware assumptions. Validate with small-scale proxy runs and current GPU pricing before committing real compute.
Special Features
- Frontier-competitive strategy (pretraining, post-training, data, systems)
- Chinchilla-optimal compute allocation (N/D/C = 6ND) with the math shown
- Loss projection and Pareto-frontier modeling from small-scale proxy runs
- GPU-hour and cloud-cost estimation with MFU-aware throughput
- Compute-optimal vs. inference-optimal (overtraining) trade-off analysis
- End-to-end roadmap: data → architecture → pretraining → alignment → eval
- Peer-reviewed foundation with explicit arXiv / lab-report citations
- Structured research hand-off package ready for an ML team
Workflow Steps
- Scope & literature review (modality, scale, baseline)
- Mathematical modeling & scaling laws (N, D, C, cost)
- Training roadmap design (data → architecture → systems)
- Post-training & alignment strategy (SFT / RLHF / DPO)
- Evaluation & safety (benchmarks, contamination, red-teaming)
- Synthesize the citation-backed hand-off package
Tags
License
MIT License — free to use, modify, and distribute. Cost and scaling estimates are planning approximations; validate against small-scale proxy runs before committing compute budgets.
Antimatter AI
antimatterai.com ↗