| name | trl-post-training |
| description | Use when building, reviewing, or editing TRL post-training workflows for agentic applications, including SFT, DPO, GRPO, RLOO, reward modeling, dataset formats, chat templates, assistant/completion-only losses, tool-calling data, reward functions, and challenge progression from SFT to environment-based RL. |
TRL Post-Training
Use this skill to design or implement post-training tasks with TRL for models
that will act as agents.
Workflow
- Identify the training stage: SFT, DPO, reward modeling, GRPO, RLOO,
environment RL, or distillation-fed retraining.
- Confirm the dataset format before choosing trainer arguments.
- Pick the smallest smoke run that exercises tokenization, generation, reward,
logging, and saving.
- Add Trackio for anything long-running or remote.
- Document the eval protocol before claiming model improvement.
Method Selection
- Use SFT first for new formats, tools, domains, and chat behavior.
- Use
$trl-sft for implementation-level SFT tasks, including trace datasets
and trl sft configs.
- Use DPO when there are high-quality chosen/rejected pairs.
- Use reward modeling when a learned scorer will be reused.
- Use GRPO when prompts can be scored by a verifier, test, parser, environment,
or judge.
- Use environment GRPO when success depends on multi-step interaction.
- Use self-distillation to recycle verified traces into later SFT or preference
data.
Implementation Rules
- Prefer TRL trainer/config classes or TRL CLI configs over custom loops.
- Use conversational
messages data for chat and tool-calling agents.
- Use prompt-only data for online RL methods such as GRPO.
- For SFT chat data, use assistant-only loss only when the chat template
supports assistant span masking.
- If
eval_strategy is enabled, provide an eval_dataset; otherwise set
evaluation to off explicitly.
- Keep reward functions deterministic where possible and log reward components.
- Keep generated checkpoints and datasets outside this context repo.
References
Read only the needed reference:
references/method-ladder.md: stage selection and challenge sequence.
references/dataset-formats.md: TRL dataset and chat-template constraints.
references/grpo-agent-rewards.md: reward functions for agentic GRPO.
references/script-patterns.md: script and config patterns.