cu-lengths-attention-flow

Bilingual guide for understanding how cu_lengths controls attention behavior across ViT and LLM stages, and how patch_positions scope differs between the two

Quality

80%

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 ./.opencode/skills/cu-lengths-attention-flow/SKILL.md

Quality

Content

77%

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

A well-sequenced, concrete, codebase-specific explanation with strong actionability and clear workflow organization, weakened mainly by bilingual duplication that inflates token count and a monolithic single-file structure that could benefit from splitting.

Suggestions

Collapse the English/Chinese duplication — e.g., present each point once and offer the translation as a single bundled reference file or a collapsible section — to recover roughly half the tokens.

Move the SFT-path detail and the quick-reference/pitfalls tables into separate reference files (e.g., SFT_PATH.md, QUICK_REFERENCE.md) and link to them from the main body to improve progressive disclosure.

Dimension	Reasoning	Score
Conciseness	The content is codebase-specific with no generic concept padding, but the entire body is duplicated in English and Chinese, roughly doubling token count, and several prose passages could be tightened — matching the 'mostly efficient but could be tightened' anchor.	2 / 3
Actionability	Provides concrete code excerpts (the cu_lengths loop, batch() branching, PackedSeqParams construction, SP padding, SFT cu_seqlens derivation) alongside exact file paths, function names, and shapes, giving specific copy-relevant guidance.	3 / 3
Workflow Clarity	The explanation is clearly sequenced — generation -> forward branching -> LLM attention behavior -> SP padding -> critical insight -> SFT path -> quick reference -> pitfalls — with no risky batch/destructive operation that would require missing validation checkpoints.	3 / 3
Progressive Disclosure	It is a single ~268-line monolithic file with no bundle references, and several inline sections (SFT path, quick-reference tables) could be split out; structure exists but content that could be separate is inline.	2 / 3
	Total	10 / 12 Passed

Description

82%

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 specific, well-targeted description with strong natural trigger terms and clear distinctiveness, but it omits an explicit 'Use when...' guidance clause, which caps completeness at 2.

Suggestions

Append an explicit trigger clause such as 'Use when reasoning about attention boundaries in the LLaVA-OneVision2 forward pass, debugging packed vs non-packed attention, or investigating cu_lengths / patch_positions behavior.' to lift completeness to 3.

Lead with the action ('Explains how...') rather than 'Bilingual guide for...' so the capability reads as a concrete third-person action.

Dimension	Reasoning	Score
Specificity	Names multiple concrete technical foci — 'cu_lengths controls attention behavior across ViT and LLM stages' and 'patch_positions scope differs between the two' — rather than vague language, matching the 'lists multiple specific concrete actions' anchor.	3 / 3
Completeness	It answers 'what' (how cu_lengths controls attention and how patch_positions scope differs) but lacks any explicit 'Use when...' trigger clause, so per the rubric guideline completeness is capped at 2.	2 / 3
Trigger Term Quality	Uses exact natural terms an engineer on this codebase would say — 'cu_lengths', 'patch_positions', 'attention behavior', 'ViT and LLM stages' — giving good coverage of the target user's vocabulary.	3 / 3
Distinctiveness Conflict Risk	The description is narrowly scoped to a distinct codebase-specific mechanism (cu_lengths/patch_positions attention control across ViT/LLM), making it unlikely to trigger for the wrong skill.	3 / 3
	Total	11 / 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
metadata_version	'metadata.version' is missing	Warning

	Total	15 / 16 Passed

Repository: EvolvingLMMs-Lab/LLaVA-OneVision-2
Path: .opencode/skills/cu-lengths-attention-flow/SKILL.md
Commit: bad6f1d

Reviewed: 4 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.