simpo-training
Simple Preference Optimization for LLM alignment. Reference-free alternative to DPO with better performance (+6.4 points on AlpacaEval 2.0). No reference model needed, more efficient than DPO. Use for preference alignment when want simpler, faster training than DPO/PPO.
- Score
- Unranked
- Evidence
- No signals yet
- Last commit
- Not known
- Listed
Install
armory install simpo-training --cli claudewrites the skill into.claude/skills/simpo-training/SKILL.mdListed as compatible
# fetches the source and writes it to:
.claude/skills/simpo-training/SKILL.mdNeeds the armory CLI · not on npm yet, build it from cli/ in the repository
What it is
Simple Preference Optimization for LLM alignment. Reference-free alternative to DPO with better performance (+6.4 points on AlpacaEval 2.0). No reference model needed, more efficient than DPO. Use for preference alignment when want simpler, faster training than DPO/PPO.
When to use it
Simple Preference Optimization for LLM alignment. Reference-free alternative to DPO with better performance (+6.4 points on AlpacaEval 2.0). No reference model needed, more efficient than DPO. Use for preference alignment when want simpler, faster training than DPO/PPO.
How to install / invoke
# Copy the skill into your .claude/skills/ directory
curl -sL https://raw.githubusercontent.com/davila7/claude-code-templates/main/cli-tool/components/skills/ai-research/post-training-simpo/SKILL.md -o .claude/skills/post-training-simpo/SKILL.md
Notes
Extracted from davila7/claude-code-templates, post-training-simpo category.