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.

Quality

81%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Low

Low-risk findings worth noting

Quality

Content

87%

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The body is a lean, highly actionable training guide with executable commands, complete configs, and a clean one-level-deep reference structure. Its one notable gap is the absence of explicit validation checkpoints in the training workflow for a costly batch operation.

Suggestions

Add a validation checkpoint before the full run, e.g., a quick smoke test on a small data slice or an instruction to confirm the loss is decreasing before committing to a full epoch.

Consolidate the three overlapping decision lists ('Use SimPO when', 'Algorithm selection', 'Use alternatives instead') into a single table to remove redundancy.

Note explicitly that scripts/run_simpo.py and accelerate_configs/deepspeed_zero3.yaml come from the cloned alignment-handbook repo, so the launch commands are unambiguous about file provenance.

Dimension	Reasoning	Score
Conciseness	The body is dense and almost entirely executable configs, commands, and troubleshooting with inline comments that earn their place (e.g., 'Critical: 3e-7 to 1e-6'), with no explanation of concepts Claude already knows; minor redundancy across the three 'when to use' lists is not enough to drop below the lean/efficient 3-anchor.	3 / 3
Actionability	Provides fully executable bash (conda/pip/accelerate launch) and complete copy-paste YAML configs with real model paths, datasets, and all hyperparameters, so it reaches the copy-paste-ready 3-anchor rather than the incomplete/pseudocode 2-anchor.	3 / 3
Workflow Clarity	The install -> config -> launch sequence is clear and troubleshooting exists, but there are no explicit validation checkpoints for an expensive batch training operation (no smoke test, no 'verify loss is decreasing before full run'); per the rubric's batch-operation rule this caps the score at 2.	2 / 3
Progressive Disclosure	The body is a concise overview with three well-signaled one-level-deep references (loss-functions/hyperparameters/datasets), each annotated with its contents, and all three files were verified to exist and be topically aligned, matching the clear-overview 3-anchor.	3 / 3
	Total	11 / 12 Passed

Description

75%

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

The description clearly states both what SimPO is and when to use it, with a distinctive reference-free DPO-alternative positioning. It is weakest on concrete-action specificity and natural trigger-term coverage, relying more on method properties and benchmark metrics than on user-voiced keywords.

Suggestions

Lead with concrete actions (e.g., 'Aligns LLMs with preference data via SimPO training') instead of leading with properties/benchmarks.

Add natural trigger variations users would say, such as 'fine-tune on preferences', 'RLHF/DPO alternative', or 'align a model with human preferences'.

Drop the 'AlpacaEval 2.0 +6.4 points' benchmark claim from the description or move it to the body; it is a performance boast, not a trigger term.

Dimension	Reasoning	Score
Specificity	Names the domain and some actions ('preference alignment', 'training') but leans on method properties ('reference-free', 'more efficient than DPO', '+6.4 points on AlpacaEval 2.0') rather than listing multiple concrete actions, so it stops short of the 3-anchor.	2 / 3
Completeness	Explicitly answers what (reference-free preference optimization, DPO alternative) and when ('Use for preference alignment when want simpler, faster training than DPO/PPO'), satisfying the explicit-trigger requirement so it is not capped at 2.	3 / 3
Trigger Term Quality	Includes usable natural terms ('preference alignment', 'DPO', 'PPO', 'LLM alignment') but omits common variations a user might actually say ('fine-tune on preferences', 'RLHF alternative', 'align a model'), and 'AlpacaEval 2.0' is a benchmark rather than a trigger.	2 / 3
Distinctiveness Conflict Risk	Occupies a clear niche (reference-free SimPO positioned against DPO/PPO) with distinct triggers, making it unlikely to fire for unrelated skills; the reference-free + DPO-alternative framing is distinctive rather than generic.	3 / 3
	Total	10 / 12 Passed

Validation

75%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 12 / 16 Passed

Validation for skill structure

Criteria	Description	Result
metadata_version	'metadata.version' is missing	Warning
metadata_field	'metadata' should map string keys to string values	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning
referenced_paths_exist	Referenced path issues: 3 missing	Warning

	Total	12 / 16 Passed

Repository: NousResearch/hermes-agent
Path: optional-skills/mlops/simpo/SKILL.md
Commit: 4c9628e

Reviewed: about 13 hours ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.