knowledge-distillation

Compress large language models using knowledge distillation from teacher to student models. Use when deploying smaller models with retained performance, transferring GPT-4 capabilities to open-source models, or reducing inference costs. Covers temperature scaling, soft targets, reverse KLD, logit distillation, and MiniLLM training strategies.

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 ./backend/cli/skills/ml-training/knowledge-distillation/SKILL.md

Quality

Content

50%

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

The content is technically rich and mostly actionable but overlong for what Claude already knows, with several non-executable code gaps and no validation checkpoints in training workflows. The unused references/minillm.md bundle file signals a missed progressive-disclosure opportunity.

Suggestions

Replace the inlined MiniLLM explanation with a one-line link to references/minillm.md and trim the toy softmax/forward-vs-reverse-KLD primers, which restate concepts Claude already knows.

Make code copy-paste executable: define or import train_dataset, optimizer, dataloader, and calculate_similarity (or remove the call), and replace Strategy 2's distill()/fine_tune() pseudocode with runnable equivalents.

Add validation checkpoints to the training workflow — e.g., evaluate the student against the teacher on a held-out set and only save/accept the checkpoint when a threshold is met, giving a validate->fix->retry feedback loop.

Dimension	Reasoning	Score
Conciseness	The ~450-line body explains concepts Claude already knows (toy softmax examples for temperature scaling, repeated forward-vs-reverse KLD explanations, and a restated loss-component breakdown), so it is 'mostly efficient but includes some unnecessary explanation' rather than the lean level-3 anchor; it is not padded enough with pure filler to score 1.	2 / 3
Actionability	Most code blocks are concrete, but several are incomplete rather than copy-paste ready: train_distilled_model() references an undefined train_dataset, calculate_similarity() is called but never defined, and Strategy 2's distill()/fine_tune() are pseudocode, fitting the level-2 'some concrete guidance but incomplete; missing key details' anchor rather than the fully-executable level-3.	2 / 3
Workflow Clarity	Steps are sequenced (install -> quick start -> strategies -> deployment), but model-training workflows have no validation checkpoints or feedback loops — no 'validate then proceed', no checkpoint/resume gate, and the Evaluation section is print statements rather than a pass/fail gate, so per the rubric's destructive/batch guidance this caps at level-2.	2 / 3
Progressive Disclosure	A dedicated bundle file references/minillm.md exists, but the body never links to it and instead inlines ~30 lines of MiniLLM content, exemplifying the level-2 anti-pattern 'content that should be separate is inline' with a reference present but not clearly signaled; it avoids level-1 only because the body itself is sectioned rather than a monolithic wall.	2 / 3
	Total	8 / 12 Passed

Description

100%

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 strong: it names concrete capabilities, provides an explicit 'Use when' trigger clause covering both what and when, and occupies a distinct niche. It does not need padding and avoids over-claims.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete actions and techniques — "Compress large language models using knowledge distillation from teacher to student models", "transferring GPT-4 capabilities to open-source models", "reducing inference costs", plus a specific technique list ("temperature scaling, soft targets, reverse KLD, logit distillation, and MiniLLM training strategies"), matching the 'lists multiple specific concrete actions' anchor.	3 / 3
Completeness	It explicitly answers both 'what' ("Compress large language models using knowledge distillation from teacher to student models") and 'when' with an explicit "Use when..." clause listing three triggers, matching the level-3 example and exceeding the level-2 anchor where 'when' is only implied.	3 / 3
Trigger Term Quality	The "Use when deploying smaller models with retained performance, transferring GPT-4 capabilities to open-source models, or reducing inference costs" clause surfaces natural phrases a user would say ('deploying smaller models', 'reducing inference costs', 'GPT-4 capabilities'), giving good coverage; it is not merely technical jargon, so it does not fall to the level-2 anchor.	3 / 3
Distinctiveness Conflict Risk	Knowledge distillation / model compression is a clear niche with distinct triggers (teacher-student, MiniLLM, reverse KLD) unlikely to collide with other skills; it is far more specific than the level-2 'Works with document files' style overlap.	3 / 3
	Total	12 / 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: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/knowledge-distillation/SKILL.md
Commit: 52845c3

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.