SFT vs DPO vs GRPO
Choose the right post-training stage from the supervision you actually have.
By the LLMix engineering practice · Updated
- Method selection
- SFT
- DPO
- GRPO
Guides
Post-Training Engineering Guides
Practical decision frameworks, implementation patterns, and failure analyses for teams customizing models and training specialized agents.
Start here
Build an RL Environment for an LLM Agent
The practical architecture: state, tools, transitions, resets, verifiers, rewards, and rollout integration.
By the LLMix engineering practice · Updated
- RL environments
- Agents
- Architecture
Why GRPO Training Fails
A triage order for runs that do not learn: rewards, variance, rollouts, KL, truncation, staleness, leakage.
By the LLMix engineering practice · Updated
- GRPO
- Failure diagnosis
- RLVR
Guides are written and maintained by LLMix with AI-assisted drafting, and revised as the underlying methods and tooling change. Updated dates change only for material revisions.
Turn the framework into a program.
LLMix scopes bespoke post-training programs around one capability: data, environments, rewards, optimization, and validated release.