cuOpt Numerical Optimization Formulation
Turn natural-language problems into clear LP, MILP, or QP mathematical formulations.
The skill is limited to optimization concepts and formulation guidance; it executes no code and requires no credentials or external access. It states QP beta status, convexity requirements, and non-convex limitations, keeping permission and external-effect risks low. Points are deducted because explicit user confirmation, liability boundaries, data-flow disclosure, and source citations in the skill body are incomplete; the supplied license metadata is NOASSERTION, leaving governance-level licensing uncertainty.
The document is internally coherent, distinguishes LP, MILP, and QP, covers several modeling patterns, and instructs the agent to ask about ambiguity. The committed evaluation report shows correctness of 67% and 37% and discoverability of 33% and 0%, indicating unstable key-path and triggering behavior. This is a static review without a reproducible test suite, so the score remains below the static ceiling of 10.
The audience, optimization domains, sentence-labeling workflow, and “no API code” boundary are reasonably clear for natural-language formulation tasks. Points are deducted for missing explicit non-fit ranges, semantic trigger rules, and Chinese-language support. The core function does not depend on overseas services, so no additional mainland-network deduction is warranted.
The materials provide a version, Apache-2.0 declaration, owner team, tags, skill card, evaluation report, and repository-level update information. Points are deducted because the static validation report identifies missing recommended Instructions and Examples sections and an invalid author format; installation/dependency guidance, FAQs, changelog, and explicit maintenance ownership are incomplete. The skill-level and repository-level licensing statements also merit clarification.
The skill contains useful patterns for sentence classification, implicit objectives, cutting stock, goal programming, inventory timing, and shared mixing. The evaluation report records Claude correctness of 67%, Codex correctness of 37%, and effectiveness results of 93% and 71%. Points are deducted because it provides concepts rather than executable models or API code, while correctness and discoverability remain inconsistent and outputs require human review.
The fixed revision, evaluation tasks, expected behaviors, result table, and static findings provide some auditability. Points are deducted because the benchmark has only three positive tasks and no negative tasks, lacks raw execution logs or independent replication, and contains no committed test suite covering the skill’s key formulations. Static review cannot independently reproduce the claims.
- Do not treat the formulations as verified solver results; manually review objectives, integrality, QP convexity, and implicit constraints.
- The evaluation sample is small and has no negative-trigger tasks; discoverability and correctness are unstable, so boundary, false-trigger, and failure-feedback tests should be added.
- Clarify the relationship between the skill license, repository license, and NOASSERTION metadata, and add explicit maintenance ownership, changelog, and authoritative citations.
What does this skill do, and when should you use it?
This skill helps users move from a problem statement to a structured formulation by identifying parameters, constraints, decisions, and objectives. It covers linear programming, mixed-integer linear programming, and quadratic programming, including variable types, convexity, sensitivity concepts, and common modeling patterns. It is conceptual and contains no cuOpt API code. When wording is ambiguous, it requires clarification or comparison of plausible interpretations.
Reads problem text and classifies each sentence or logical clause as a parameter, constraint, decision, or objective; detects implicit constraints and objectives; summarizes the model elements before formulation; explains when LP, MILP, or QP is appropriate; and presents formulation patterns for piecewise-linear objectives, cutting stock, goal programming, multi-period inventory, and shared mixing.
- An operations analyst turns production, resource, and cost descriptions into an LP or MILP.
- A scheduling, facility-location, or selection problem includes integer or binary decisions.
- A financial analyst formulates portfolio variance with return and budget constraints as a QP.
- A user needs to resolve timing assumptions in a purchasing or inventory problem before modeling.
- A modeler reviews variable types and objective construction in cutting-stock, goal-programming, or shared-mixing formulations.
What are this skill's strengths and limitations?
- Covers core formulation concepts for LP, MILP, and QP.
- Requires sentence-level parsing and explicit identification of implicit objectives and constraints.
- Includes several practical formulation patterns and modeling pitfalls.
- Documents QP minimization, convexity, and dual/sensitivity limitations.
- Provides no cuOpt API code.
- QP support is identified as beta in the source material.
- Ambiguous problems require clarification or multiple interpretations.
- The SKILL.md provides no test suite or concrete execution-validation evidence.
How do you install this skill?
Install the specific skill from NVIDIA’s collection with:
npx skills add nvidia/skills --skill cuopt-numerical-optimization-formulation --yes
The README does not document the exact installation directory; the CLI prompts for the destination.
How do you use this skill?
After the skill is loaded by an Agent, provide a formulation task such as: “Turn this production-planning problem into an LP or MILP. First classify the parameters, constraints, decisions, and objective, then write the mathematical formulation.” The source does not document client-specific triggering or runtime procedures.