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

75%

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

Fix and improve this skill with Tessl

tessl review fix ./backend/cli/skills/ml-training/simpo/SKILL.md

Quality

Content

80%Weight 40%Scale 1-5

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

The body is highly actionable with complete, copy-paste-ready configs and commands, and it uses progressive disclosure well by pushing detail into real reference files. Its main weakness is the absence of an explicit validation/feedback checkpoint in the training workflow.

Suggestions

Add a validation step to each training workflow, e.g. after launch: 'Watch the first ~50 steps; if loss diverges, stop and apply the Common issues: Loss divergence remedies before restarting.'

De-duplicate the accelerate launch command between Quick start and Workflow 1 by referencing it once and varying only the config path.

Tighten the 'When to use vs alternatives' prose into a compact table to recover tokens.

Dimension	Reasoning	Score
Conciseness	Mostly lean and practical (configs, commands, troubleshooting deltas) with little concept re-explanation, but the Quick start launch command is duplicated in Workflow 1 and some inline commentary could be trimmed.	4 / 5
Actionability	Fully copy-paste ready: complete conda/pip install steps, accelerate launch commands, and complete YAML configs with real model paths and datasets across three model workflows, plus concrete config-change remedies for each issue.	5 / 5
Workflow Clarity	The install -> config -> launch sequence is clear, but the workflow has no validation checkpoint (e.g. 'monitor loss; if diverging, reduce LR') before committing to a long training run; per the batch-operation cap this stays at 3 even though troubleshooting is covered reactively in Common issues.	3 / 5
Progressive Disclosure	SKILL.md is a concise overview with a dedicated Advanced topics section linking to three real, one-level-deep reference files (loss-functions.md, hyperparameters.md, datasets.md), each clearly signaled with markdown links and appropriately scoped.	5 / 5
	Total	17 / 20 Passed

Description

70%Weight 40%Scale 1-5

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 is third-person, specific, and includes both a clear 'what' and an explicit 'Use for...' trigger clause. It is slightly benefit-heavy rather than action-comprehensive, and the trigger phrasing is narrower than ideal.

Suggestions

Add a few concrete trigger synonyms a user might say, e.g. 'RLHF alternative' or 'preference fine-tuning', to broaden natural keyword coverage.

Reframe benefits as discrete actions (e.g. 'Trains a policy on chosen/rejected pairs without a reference model') to lift specificity.

Make the 'when' clause enumerate concrete scenarios rather than a single condition, e.g. 'Use when aligning a chat or instruct model on preference data and you want a reference-free, compute-efficient alternative to DPO/PPO.'

Dimension	Reasoning	Score
Specificity	Names the domain ('Simple Preference Optimization for LLM alignment', 'preference alignment') and 1-2 concrete actions (reference-free training, no reference model), but mostly lists benefits rather than a comprehensive set of discrete actions.	3 / 5
Completeness	Clearly states what the skill does ('reference-free alternative to DPO') and includes an explicit 'Use for preference alignment when want simpler, faster training than DPO/PPO' trigger, though the when-clause is somewhat narrow rather than enumerating concrete trigger scenarios.	4 / 5
Trigger Term Quality	Good keyword coverage with natural practitioner terms ('DPO', 'PPO', 'preference alignment', 'LLM alignment', 'reference-free', 'SimPO'); a few common synonyms like 'RLHF' or 'reward model' are missing.	4 / 5
Distinctiveness Conflict Risk	SimPO is a named, niche algorithm with distinct reference-free triggers, but it carries minor overlap risk with a closely related DPO/preference-training skill.	4 / 5
	Total	15 / 20 Passed

Validation

87%

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 — 14 / 16 Passed

Validation for skill structure

Criteria	Description	Result
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	14 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/simpo/SKILL.md
Commit: fcd2e11

Reviewed: 2 days 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.