Training Agents

Public Codex context for agentic post-training work with TRL.

This repository contains reusable instructions, sub-agent definitions, skills, and lightweight guides for planning, implementing, reviewing, and monitoring agent training workflows.

It is not a training codebase. Keep checkpoints, datasets, logs, and experiment outputs outside the tracked repo, usually under ignored workspaces/ directories or separate project repositories.

Examples

• examples/gemma4-pi-mono-sft/: TRL SFT example for google/gemma-4-E2B-it on badlogicgames/pi-mono, with Hugging Face Jobs, LoRA, hosted Trackio logging, verified Job IDs, Inspect AI HumanEval/MBPP coding evals, and private adapter artifact repos.

Guides

• program.md: operating model for Training Agents.
• docs/program.md: staged challenge ladder from SFT to environment GRPO and self-distillation.
• docs/looping-rl.md: blog post on loop-shaped reinforcement learning for agent training systems.
• docs/terminal-bench-loop.md: loop-shaped automation contract for training an approximately 2B open model toward Terminal-Bench performance above 40.