CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

llava-onevision2-consistency

Bilingual guide for running and interpreting LLaVA-OneVision2 HF vs Megatron consistency checks across TP and PP settings

Quality

76%

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

Fix and improve this skill with Tessl

tessl review fix ./.opencode/skills/llava-onevision2-consistency/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

77%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 exceptionally actionable with concrete commands, thresholds, failure diagnostics, and validation checkpoints — a model workflow-clarity document. Its weaknesses are structural: the full English/Chinese duplication doubles token cost, and the long single-file format inlines reference-grade detail (weight mappings, fixture docs) that should be offloaded to bundle files.

Suggestions

Eliminate wholesale bilingual duplication: keep one primary language and translate only the critical warnings/decision points, cutting roughly half the tokens while preserving clarity.

Move the detailed reference tables (key weight mapping, fixture descriptions, per-test internals) into files under references/ (e.g., references/weight-mapping.md, references/fixtures.md), keeping SKILL.md as an overview with clearly signaled one-level-deep links.

Trim the per-test breakdowns to a one-line purpose plus threshold each, and push failure-cause details for each test into the troubleshooting reference file.

DimensionReasoningScore

Conciseness

Per-language the body is dense and free of concept explanations Claude already knows, but every single statement is duplicated verbatim in English and Chinese (~370 lines where ~185 would suffice), which is systematic padding. This sits between 'noticeably verbose; several unnecessary explanations or padded sections' (2) and 'efficient' (4): mostly efficient content, but the wholesale bilingual mirroring is a clear tightening opportunity.

3 / 5

Actionability

Fully executable throughout: copy-paste commands ('bash tests/consistency/run_consistency_tests.sh', 'docker exec -it llava_megatron_container_ax bash', pytest invocations with -k filters), a complete env-var table with defaults, exact cosine thresholds, a symptom→cause→fix table, and a concrete weight-mapping table. This matches 'Fully executable; copy-paste ready code or commands' and exceeds the 4 anchor, which allows minor gaps.

5 / 5

Workflow Clarity

Multi-step processes are explicitly sequenced with validation checkpoints: a 7-step ordered troubleshooting checklist, a diagnosis priority order ('If this fails, all other tests are unreliable'), 'What run_consistency_tests.sh does' as a numbered 5-step sequence, and clear feedback loops ('Fix this first before investigating other tests', 'Run fast tests first'). This matches the top anchor 'Clear sequence with explicit validation steps; feedback loops for error recovery; checklists'.

5 / 5

Progressive Disclosure

No bundle files exist (references/, scripts/, assets/ are all absent) and the document is a ~370-line single file inlining detailed reference material (per-test weight-mapping tables, per-fixture documentation, known-lessons) that clearly belongs in separate reference files. Section headers are well-organized, which keeps it above the 2 anchor's 'minimal structure', but it matches 'content that should be separate is inline' rather than the 4 anchor's 'most content is appropriately placed'.

3 / 5

Total

16

/

20

Passed

Description

65%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 highly specific to its niche and low-conflict, but it omits any 'when to use' trigger guidance, which caps its completeness, and it lists only two actions despite the skill covering conversion, re-sharding, and failure diagnosis. Adding an explicit 'Use when...' clause would materially raise the score.

Suggestions

Add an explicit trigger clause, e.g., 'Use when validating that HuggingFace and Megatron/MCore checkpoints of LLaVA-OneVision2 behave identically, or when changing TP/PP sharding without retraining.'

Broaden the action list beyond 'running and interpreting' to mention the concrete capabilities documented in the body: auto HF→mcore conversion, TP/PP-customized runs, mcore→HF re-sharding, and failure diagnosis.

Include one or two natural synonyms users might say, such as 'checkpoint parity' or 'model equivalence', to strengthen trigger coverage.

DimensionReasoningScore

Specificity

The description names a precise domain ("LLaVA-OneVision2 HF vs Megatron consistency checks across TP and PP settings") with two concrete actions ("running and interpreting"), but stops there — no mention of conversion, re-sharding, or failure diagnosis. This matches the anchor 'Names domain and 1-2 concrete actions, but not comprehensive'; a 4 would require several listed specific actions.

3 / 5

Completeness

The 'what' is clear (bilingual guide for running and interpreting consistency checks), but there is no 'Use when...' clause or equivalent trigger guidance anywhere in the description, which caps completeness at 3 per the rubric guidelines. It is not a 2 because the 'what' half is concrete and specific.

3 / 5

Trigger Term Quality

Domain-natural terms like "consistency checks", "HF vs Megatron", "TP and PP" are exactly what a user in this repo would say. A 5 would require synonym coverage (e.g., checkpoint parity, model equivalence, convert); a 3 would mean common variations are broadly missing, which is not the case.

4 / 5

Distinctiveness Conflict Risk

The description is anchored to a single framework ("LLaVA-OneVision2") and a single task (HF vs Megatron consistency), a clear niche with minimal overlap risk against generic testing or conversion skills. It clearly fits the top anchor 'Clear niche with distinct triggers; minimal conflict risk'.

5 / 5

Total

15

/

20

Passed

Validation

93%

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

CriteriaDescriptionResult

metadata_version

'metadata.version' is missing

Warning

Total

15

/

16

Passed

Repository

EvolvingLMMs-Lab/LLaVA-OneVision-2

Path

.opencode/skills/llava-onevision2-consistency/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.