nlp-alignment

Best practices for LLM alignment techniques including RLHF, DPO, and instruction tuning. Use when working on alignment or safety.

Quality

75%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./researchclaw/skills/builtin/domain/nlp-alignment/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.

A tight, well-structured alignment reference with concrete hyperparameters and a clear training sequence. Its weakest area is workflow clarity, since the recipe lacks explicit validation/feedback checkpoints for the multi-step training process.

Suggestions

Add explicit validation checkpoints to the training recipe (e.g., 'After SFT, evaluate on held-out instructions before proceeding to preference optimization').

Tighten under-specified steps like 'Start with SFT on high-quality instruction data' by noting dataset size/quality criteria.

Optionally include a minimal executable snippet (e.g., a TRL config) to move actionability from concrete guidance to copy-paste ready.

Dimension	Reasoning	Score
Conciseness	The body is lean and bullet-driven, lists concrete hyperparameters without over-explaining concepts Claude already knows, and nearly every token earns its place (only trivial padding like '(complex but powerful)').	5 / 5
Actionability	Provides concrete, specific guidance such as exact hyperparameters (DPO lr=5e-7, beta=0.1; PPO lr=1e-6, clip=0.2, KL coeff=0.02) and named evaluation benchmarks, but has no executable code or commands and a couple of steps ('Start with SFT on high-quality instruction data') are slightly underspecified.	4 / 5
Workflow Clarity	The 'Training recipe' gives a clear sequence (SFT then DPO/PPO then evaluate) but lacks explicit validation checkpoints or feedback loops for error recovery, leaving checkpoints implicit.	3 / 5
Progressive Disclosure	A single under-50-line file with three well-organized sections (Methods, Training recipe, Common pitfalls) and no need for external references, meeting the simple-skill exception for full marks.	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.

A solid description that covers the what and when with concrete technique names and an explicit use-when clause. Its main limitation is framing capabilities as topics rather than actions and a somewhat narrow trigger list.

Suggestions

Reframe the 'what' as concrete actions (e.g., 'Implements and tunes RLHF, DPO, and instruction-tuning pipelines for LLM alignment') to lift specificity.

Broaden the 'when' clause with more trigger phrases and synonyms (e.g., 'Use when working on alignment, preference learning, reward modeling, or safety training').

Narrow the 'safety' trigger to reduce overlap with general safety skills, e.g., 'alignment-related safety' or 'safety fine-tuning'.

Dimension	Reasoning	Score
Specificity	Names the domain ('LLM alignment') and enumerates specific techniques (RLHF, DPO, instruction tuning), but frames them as topics ('Best practices for...') rather than concrete actions, so it does not reach the comprehensive multi-action level.	3 / 5
Completeness	Explicitly answers both 'what' ('Best practices for LLM alignment techniques including RLHF, DPO, and instruction tuning') and 'when' ('Use when working on alignment or safety'), but the when clause is limited to two trigger terms and could be more specific.	4 / 5
Trigger Term Quality	Includes natural terms a practitioner would say ('alignment', 'RLHF', 'DPO', 'instruction tuning', 'safety') with an explicit 'Use when working on alignment or safety' trigger, though a few common synonyms like 'preference learning' or 'reward model' are absent.	4 / 5
Distinctiveness Conflict Risk	Targets a fairly distinct niche (alignment/preference training) with technique-specific triggers, though the broad term 'safety' creates minor overlap risk with general safety skills.	4 / 5
	Total	15 / 20 Passed

Validation

100%

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 — 16 / 16 Passed

Validation for skill structure

No warnings or errors.

Repository: aiming-lab/AutoResearchClaw
Path: researchclaw/skills/builtin/domain/nlp-alignment/SKILL.md
Commit: be4ba47

Reviewed: 5 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.