CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

emerging-techniques-model-merging

Merge multiple fine-tuned models using mergekit to combine capabilities without retraining. Use when creating specialized models by blending domain-specific expertise (math + coding + chat), improving performance beyond single models, or experimenting rapidly with model variants. Covers SLERP, TIES-Merging, DARE, Task Arithmetic, linear merging, and production deployment strategies.

Quality

82%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

SKILL.md
Quality
Evals
Security

Quality

Content

67%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 thorough, highly actionable reference with copy-paste mergekit configs for every merge method, sensible best practices, and a real evaluation/deployment path. The main weaknesses are duplicated examples between sections (SLERP config, compatibility guidance, method selection), a few non-executable commands, and overlap between the inline method guide and references/methods.md.

Suggestions

De-duplicate the repeated examples: keep one basic SLERP config (Quick Start) and one compatibility warning (Best Practices), removing the near-verbatim repeats in the Merge Methods Guide and Common Pitfalls.

Fix or contextualize non-executable commands: replace 'python -m transformers.models.auto ...' with a working transformers loading snippet, and identify where convert.py / quantize_gptq.py come from (e.g. llama.cpp).

Add an explicit ordered workflow (choose method → write config → merge → benchmark → deploy) that ties the density-tuning and benchmark-before-deploy checkpoints into one sequence, and trim the inline Core Concepts/Merge Methods detail in favor of references/methods.md.

DimensionReasoningScore

Conciseness

The body is mostly concrete configuration examples, but there is real duplication: the basic SLERP config appears nearly verbatim in both Quick Start and the Merge Methods Guide, model-compatibility examples appear in both Best Practices #1 and Pitfall 1, and the method-selection guidance in Best Practices #3 repeats the 'Best for' lines of the guide. This fits the score-3 anchor ('could be tightened') rather than score 2, since the bulk is actionable content rather than padded prose explaining concepts Claude already knows.

3 / 5

Actionability

Most guidance is copy-paste ready: complete mergekit YAML configs for every method, install commands, and 'mergekit-yaml config.yml ./merged-model --cuda' run commands. It stops short of score 5 because some commands are not executable as written ('python -m transformers.models.auto --model_name_or_path ./merged-model' is not a real invocation, and 'convert.py'/'quantize_gptq.py' are used without identifying their source repos).

4 / 5

Workflow Clarity

The implicit end-to-end flow (install → choose method → write config → run merge → evaluate → deploy) is present and reinforced with checkpoints: the density-tuning section gives an explicit feedback loop ('If performance degrades, reduce sparsity') and Pitfall 3 mandates benchmarking before deployment. It does not reach score 5 because the sequence is never consolidated into an explicit ordered workflow with per-step validation.

4 / 5

Progressive Disclosure

The bundle is well structured: references/methods.md, references/examples.md, and references/evaluation.md are real files, referenced one level deep in a 'See Also' section with one-line descriptions. It falls short of score 5 because the ~200-line inline Core Concepts and Merge Methods Guide sections duplicate material that belongs in references/methods.md, so content placement could be tighter.

4 / 5

Total

15

/

20

Passed

Description

92%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 strong description: it states concrete capabilities, enumerates the supported merge methods and deployment, and provides an explicit 'Use when...' clause with natural trigger scenarios in third-person voice. The only gap is a few missing synonyms such as 'combine models' or 'model soup' that users might naturally say.

Suggestions

Add one or two natural synonyms users might say, e.g. 'combining models' or 'model soup', to widen trigger coverage.

Consider mentioning the mergekit YAML config workflow ('write a merge config, run mergekit-yaml') as a concrete action phrase users may search for.

DimensionReasoningScore

Specificity

The description lists multiple concrete actions ('Merge multiple fine-tuned models using mergekit to combine capabilities without retraining', 'blending domain-specific expertise (math + coding + chat)') and comprehensively names the covered methods (SLERP, TIES-Merging, DARE, Task Arithmetic, linear merging, production deployment). It goes beyond the score-4 anchor, whose coverage has minor gaps, since both actions and method coverage are complete.

5 / 5

Completeness

It explicitly answers both 'what' ('Merge multiple fine-tuned models using mergekit... Covers SLERP, TIES-Merging, DARE, Task Arithmetic, linear merging, and production deployment strategies') and 'when' ('Use when creating specialized models by blending domain-specific expertise (math + coding + chat), improving performance beyond single models, or experimenting rapidly with model variants') with concrete trigger phrases, matching the top anchor exactly.

5 / 5

Trigger Term Quality

Natural terms like 'merge', 'mergekit', 'fine-tuned models', 'blending', and specific method names (SLERP, TIES, DARE) are present and would be said by users needing this skill. It falls short of the score-5 anchor because common variations such as 'combine models', 'model soup', or 'fuse models' are missing.

4 / 5

Distinctiveness Conflict Risk

Model merging via mergekit is a clear niche with distinct triggers (merge method names, mergekit, blending fine-tuned models), making confusion with adjacent skills like fine-tuning or evaluation unlikely. It is not scored lower because no meaningful overlap with sibling skills is evident.

5 / 5

Total

19

/

20

Passed

Validation

87%

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

CriteriaDescriptionResult

skill_md_line_count

SKILL.md is long (541 lines); consider splitting into references/ and linking

Warning

metadata_field

'metadata' should map string keys to string values

Warning

Total

14

/

16

Passed

Repository

davila7/claude-code-templates

Path

cli-tool/components/skills/ai-research/emerging-techniques-model-merging/SKILL.md

Reviewed

Table of Contents

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.