rwkv-architecture

RNN+Transformer hybrid with O(n) inference. Linear time, infinite context, no KV cache. Train like GPT (parallel), infer like RNN (sequential). Linux Foundation AI project. Production at Windows, Office, NeMo. RWKV-7 (March 2025). Models up to 14B parameters.

Quality

59%

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/rwkv/SKILL.md

Quality

Content

65%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 well-structured, largely executable reference: code examples dominate, deep architecture is correctly deferred to real reference files, and troubleshooting covers common failure modes. Main weaknesses are padded educational comparison content that Claude doesn't need, and workflows lacking any validation checkpoints.

Suggestions

Collapse the Workflow 4 speed-comparison block and the KV-cache arithmetic into a single short comparison table — the token-by-token explanation of quadratic attention is content Claude already knows.

Make the fine-tuning workflow self-contained: drop the unused `RWKVTrainer` import, show how `train_dataloader` is constructed, and use one consistent model initialization style.

Add validation checkpoints to the workflows, e.g. sanity-check logits before the streaming loop, verify state tensor shapes after chunked processing, and check training loss before/after fine-tuning epochs.

Dimension	Reasoning	Score
Conciseness	Mostly code-driven and lean (install, basic usage, state-passing wrong/right pattern), but several sections pad concepts Claude already knows: the token-by-token speed comparison ('First token: 1 computation / Second token: 2 computations / ... 1000th token: 1000 computations') re-explains quadratic attention, the KV-cache arithmetic in Workflow 4 is tutorial-style, and 'Key advantages' repeats the description frontmatter verbatim. Not score 4 because these padded blocks are more than trimmable one-liners — two full subsections are expendable.	3 / 5
Actionability	Mostly executable, copy-paste-ready guidance: install commands with pinned versions, working GPT/RNN-mode forward calls, streaming generation with PIPELINE, and concrete issue→fix troubleshooting snippets. Not score 5 because Workflow 2 leaves `load_document()`/`chunks()` undefined, Workflow 3 imports `RWKVTrainer` without using it, references an undefined `train_dataloader`, and mixes `RWKV(config)` with the `RWKV(model=...)` initialization used elsewhere; these are minor but real gaps.	4 / 5
Workflow Clarity	The four workflows are clearly organized by scenario with code showing the operation order, but they are recipes rather than sequenced steps, and validation checkpoints are absent throughout — no check that generation logits look sane, no loss/eval verification in fine-tuning, no state-shape check in streaming. Anchor 3 ('sequence present but checkpoints missing or implicit') fits; not score 4 because no workflow demonstrates a validate-then-proceed step, though no destructive/batch cap applies.	3 / 5
Progressive Disclosure	The body is a well-organized overview (quick start, workflows, when-to-use, troubleshooting, hardware) that pushes deep material to three real, one-level-deep reference files, each clearly signaled with a preview of its contents ('See [references/architecture-details.md](references/architecture-details.md) for WKV operation, time-decay mechanism, and receptance gates'); all three referenced files exist in the bundle. This matches the anchor-5 example structure of quick start plus well-signaled references with no nesting.	5 / 5
	Total	15 / 20 Passed

Description

53%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 information-dense and factually specific, but it describes the RWKV architecture rather than the skill's capabilities, and it completely lacks a 'Use when...' trigger clause. The most natural trigger term ('RWKV' unqualified) is also absent, weakening discovery.

Suggestions

Add an explicit 'Use when...' clause, e.g. 'Use when working with RWKV models, linear-attention architectures, or constant-memory long-context inference.'

State the skill's concrete actions (e.g., 'Train, fine-tune, and run inference with RWKV models') instead of only reciting architecture properties.

Include the natural keyword 'RWKV' itself plus synonyms such as 'linear attention' and 'constant-memory inference' for better trigger matching.

Dimension	Reasoning	Score
Specificity	The description names the domain with several concrete technical attributes ('O(n) inference', 'no KV cache', 'Train like GPT (parallel), infer like RNN (sequential)') but lists no actions the skill actually performs — it reads as an architecture fact sheet, not a capability list. It is not score 2 because the characteristics are specific rather than generic, and not score 4 because 'Lists several specific actions' is unmet — there are no skill actions at all (no training, fine-tuning, or inference verbs).	3 / 5
Completeness	The 'what' is present in concrete terms (a linear-time RNN+Transformer hybrid architecture), but there is no 'Use when...' clause or any equivalent trigger guidance, capping completeness at 3 per the rubric guideline. It is not score 2 because the 'what' is clear and information-dense rather than vague.	3 / 5
Trigger Term Quality	It contains some relevant keywords ('RNN+Transformer hybrid', 'KV cache', 'infinite context', 'RWKV-7') but the standalone term 'RWKV' itself — the most natural thing a user would say — never appears on its own, and phrases skew technical (O(n), NeMo) with no synonyms like 'linear attention' or 'constant memory'. Not score 4 because the primary natural trigger term is effectively missing and common variations are absent.	3 / 5
Distinctiveness Conflict Risk	'RWKV-7 (March 2025)', 'Linux Foundation AI project', and 'Production at Windows, Office, NeMo' carve out a distinct niche with minimal conflict risk. Not score 5 because leading with 'RNN+Transformer hybrid' creates minor overlap risk with general RNN, Transformer, or state-space-model (Mamba) skills before the RWKV-specific tokens appear.	4 / 5
	Total	13 / 20 Passed

Validation

87%

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 — 14 / 16 Passed

Validation for skill structure

Criteria	Description	Result
metadata_version	'metadata.version' is missing	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	14 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/rwkv/SKILL.md
Commit: 4082a2e

Reviewed: 1 day 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.