Post-training, organized for your harness.
Post-training Skills is a reusable Skills.md artifact and a local deep-agent harness for teams that want to design, run, and improve LLM post-training in their own stack.
A simple walkthrough
Start with the artifact
Open Skills.md and use it as the operating map for a post-training program. It moves from problem framing and governed data through SFT, preference collection, reward modeling, RLHF/RLAIF/RLVR, DPO-family optimization, safety alignment, tool-use trajectories, evaluation, inference constraints, deployment, and feedback loops.
Adapt each skill to your system
Every module is written around decisions and deliverables: purpose, prerequisites, inputs, process, outputs, quality gates, tools, failure modes, safety notes, and a practice task. Replace the examples with your datasets, base model, policies, evaluators, serving runtime, and release requirements.
Use the agent when you need a revision
The local agent uses LangChain deepagents. Backboard.io provides the Luna 5.6 model, Tavily provides current web research, and focused subagents separate research, source review, curriculum architecture, implementation writing, and critique. The result is written back to Skills.md so the artifact remains the source of truth.
Keep quality gates close to the work
Post-training is not complete when a checkpoint trains. Compare against a baseline, check capability and safety slices, inspect preference and reward behavior, test tool calls, measure latency and cost on the real serving path, red-team the system, document known limitations, and keep a rollbackable release artifact.
Download or scaffold
For the guide alone, download the Markdown with npx. To run the builder locally, scaffold the repository, install the Python package, add your Backboard and Tavily keys, and invoke the agent with a focused request.
Guide only
npx github:Auro-rium/posttraining-skills --out Skills.md
Full local harness
npx github:Auro-rium/posttraining-skills --init "posttraining skills" cd "posttraining skills" cp .env.example .env # add BACKBOARD_API_KEY and TAVILY_API_KEY python3 -m venv .venv source .venv/bin/activate pip install -e . posttraining-skills-agent
What this is—and is not
This is not a hosted training platform or a claim that one provider is the right answer. It is a portable set of decisions, artifacts, and checks that can sit beside your existing training and evaluation harness. Use the guide to make your own workflow explicit, then verify provider behavior, data policy, model quality, and safety requirements before shipping.
The current artifact was generated with Backboard Luna 5.6 and Tavily research, then kept in git so changes remain inspectable and reproducible.