fine-tuning-with-trl

Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training. Use when need RLHF, align model with preferences, or train from human feedback. Works with HuggingFace Transformers.

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 ./skills/post-training/trl-fine-tuning/SKILL.md

Quality

Content

72%

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

A highly actionable skill body with copy-paste executable examples for every major TRL method and clean one-level-deep progressive disclosure to four real reference files. The main weaknesses are workflow validation gaps (training pipelines lack explicit verify/recovery checkpoints) and minor verbosity in section intros.

Suggestions

Add explicit validation checkpoints to the multi-step training workflows, e.g. after SFT a 'verify training loss decreased and save exists' step, and before PPO a 'validate reward model accuracy on a held-out preference set' step.

Tighten or remove restating section intros like 'Complete pipeline from base model to human-aligned model' and 'Align model with preferences without reward model' since the checklists and code already convey the purpose.

Replace the final Evaluate step's single-generation sample with a concrete quantitative check (e.g. compare reward or win-rate against the SFT baseline) so the feedback loop is actionable.

Dimension	Reasoning	Score
Conciseness	The body is mostly efficient with executable snippets and checklists, but includes some restating of what each method is (e.g. 'Complete pipeline from base model to human-aligned model', 'Align model with preferences without reward model') and slightly padded section intros that could be tightened. It assumes Claude's competence but is not uniformly lean.	2 / 3
Actionability	Provides fully executable, copy-paste-ready Python and CLI examples for SFT, DPO, PPO, GRPO, and reward modeling with concrete model names, hyperparameters, and dataset references, matching the anchor for executable code and specific examples.	3 / 3
Workflow Clarity	Multi-step workflows are sequenced with copy-this checklists (RLHF pipeline, DPO, GRPO), but they lack explicit validation/verification checkpoints — e.g. the RLHF pipeline has no 'verify SFT loss converged' or 'evaluate reward model accuracy before PPO' step, and the evaluate step is only a generation sample. Per the rubric, missing validation steps in batch training operations caps this at 2.	2 / 3
Progressive Disclosure	The SKILL.md is an overview with four clearly signaled, one-level-deep references (sft-training.md, dpo-variants.md, reward-modeling.md, online-rl.md), all real files with no nested references, and content is appropriately split with easy navigation via the Advanced topics section.	3 / 3
	Total	10 / 12 Passed

Description

77%

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 strong, specific description that clearly enumerates four concrete training methods and provides an explicit 'Use when' trigger clause covering RLHF, preference alignment, and human feedback. Weaker on trigger-term coverage and distinctiveness, due to slightly clipped phrasing and a broad closing line.

Suggestions

Expand trigger terms to include natural variations users would say, e.g. 'preference data', 'chosen/rejected pairs', 'instruction tuning', and 'reward model' rather than only RLHF/human feedback.

Fix the grammatically clipped 'Use when need RLHF' to a full clause such as 'Use when you need RLHF' for natural readability.

Sharpen distinctiveness by replacing the generic 'Works with HuggingFace Transformers' with a more niche-scoping phrase that separates TRL from plain Trainer/PEFT fine-tuning skills.

Dimension	Reasoning	Score
Specificity	Names multiple concrete actions: 'SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training' — a comprehensive list of specific capabilities, matching the anchor that lists multiple specific concrete actions.	3 / 3
Completeness	Explicitly answers both 'what' (the four fine-tuning methods listed) and 'when' via the 'Use when need RLHF, align model with preferences, or train from human feedback' trigger clause, satisfying the anchor requiring both an explicit 'what' and an explicit trigger.	3 / 3
Trigger Term Quality	Includes some natural terms users would say ('RLHF', 'align model with preferences', 'train from human feedback') but misses common variations like 'preference data', 'instruction tuning', 'reward model', and the phrasing 'Use when need RLHF' is grammatically clipped. Some relevant keywords present but coverage is incomplete.	2 / 3
Distinctiveness Conflict Risk	The TRL/RLHF niche is fairly distinct from generic fine-tuning skills, but the closing 'Works with HuggingFace Transformers' is broad and the description could overlap with general fine-tuning or PEFT skills; it does not fully carve out a conflict-free niche.	2 / 3
	Total	10 / 12 Passed

Validation

93%

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 — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

Repository: OpenLAIR/dr-claw
Path: skills/post-training/trl-fine-tuning/SKILL.md
Commit: 6da7f7c

Reviewed: about 8 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.