Content
60%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body delivers strong, executable coverage of installation, model usage, pretrained loading, Mamba-1 vs Mamba-2 differences, and benchmarking, with a useful troubleshooting section. Its two real defects are broken progressive disclosure — most reference links are dead while real bundle files sit orphaned — and repetitive performance claims that inflate the token budget without adding information.
Suggestions
Fix the reference links to match the actual bundle: point Advanced topics at `references/architecture-details.md` (selective SSM math), `references/benchmarks.md` (performance), and a Mamba-2 section within the existing files, and add a link to `references/training-guide.md` so it is discoverable.
Deduplicate the "5× faster / no KV cache / linear scaling" claims into a single Performance section and remove the star count and paper dates or move them to a clearly dated Resources note.
Quote the version pin (`pip install "causal-conv1d>=1.4.0"`) so the install command is actually copy-paste executable, and either include the benchmark script in the bundle or note it must be run from the mamba GitHub repo.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly efficient code-first content, but the "5× faster inference" / "No KV cache" / "linear vs quadratic" claims are repeated across Quick start, Workflow 4 expected results, Advantages, Hardware requirements, and Performance sections, and time-sensitive details (paper dates, "⭐ 13,000+" stars) pad the Resources section — anchor 3, could be tightened but not severely verbose. | 3 / 5 |
Actionability | Mostly executable, copy-paste-ready guidance: install commands, Mamba block usage, full LM config with generation, pretrained-model loading, and troubleshooting fixes. It falls short of anchor 5 because `pip install causal-conv1d>=1.4.0` is unquoted (bash treats `>=` as a redirect, so the command fails as written) and the benchmark workflow invokes `benchmarks/benchmark_generation_mamba_simple.py`, a path that does not exist in the skill bundle. | 4 / 5 |
Workflow Clarity | Workflows 1–4 are clearly sequenced (install → basic block → LM → pretrained models → comparison → benchmarking) and the Common issues section provides error-recovery guidance (OOM → gradient checkpointing, load failure → use `MambaLMHeadModel.from_pretrained`). No explicit validation checkpoints exist, but the operations are non-destructive inference tasks, so this is a minor gap — anchor 4 rather than 5. | 4 / 5 |
Progressive Disclosure | Three of the four Advanced-topics links point to files that do not exist (`references/selective-ssm.md`, `references/mamba2-details.md`, `references/performance.md`), while the actual bundle files (`architecture-details.md`, `benchmarks.md`, `training-guide.md`) are never referenced from SKILL.md, and `rwkv/overview.md` chains into a further nested `references/rwkv/references/` level. Navigation to the bundle is factually broken, which is worse than the "not clearly signaled" state of anchor 3 and fits anchor 2's minimal effective structure. | 2 / 5 |
Total | 13 / 20 Passed |