Sentence classification
| Sentence | Label | Note |
|---|---|---|
| "Operates 2 ovens" | Parameter | Both ovens are available. This is not an oven-selection problem. |
| "Plans to bake 300 loaves this week" | Constraint (implicit) | Committed phrasing, so bread = 300 exactly. |
| "May also bake cakes" | Decision | The number of cakes is chosen. |
| "Cakes earn $12, cost $5; bread earns $4, costs $1.50" | Parameters | Net per unit is $7 per cake and $2.50 per loaf. |
| "Each oven runs at most 40 hours" | Parameter (40) and constraint | Oven time used is at most 2 × 40 = 80 hours. |
| "A loaf takes 0.1 oven-hours, a cake 0.4" | Parameters | They set the oven-time usage. |
| "Determine the weekly baking plan" | Objective (implicit) | Revenue and costs are both given, so maximize total profit. |
Formulation
- Decision: c = number of cakes (≥ 0). Integer in practice, though the LP optimum is already whole.
- Fixed by commitment: loaves b = 300.
- Objective: maximize 2.5·b + 7·c, which is profit = revenue − ingredient cost.
- Constraint: 0.1·b + 0.4·c ≤ 80 oven-hours.
- Problem type: LP, or a trivial MILP if cakes must be integers.
Solution
- Bread uses 300 × 0.1 = 30 oven-hours, leaving 50 hours.
- Each cake nets a positive $7, so the oven time left over goes to cakes: c = 50 / 0.4 = 125 cakes.
- The plan splits across the two ovens without trouble. For example, oven 1 bakes all 300 loaves and 25 cakes (30 + 10 = 40 h), and oven 2 bakes 100 cakes (40 h).
Plan: 300 loaves and 125 cakes. Profit = 300 × 2.50 + 125 × 7 = $750 + $875 = $1,625. Revenue is $2,700 and ingredient cost is $1,075. Both ovens are fully used at 80 of 80 hours.
Ambiguities
- "Plans to bake 300 loaves": I read this as exactly 300. If it meant at least 300, the answer changes. Bread earns $25 per oven-hour against $17.50 for cakes, so the plan would become 800 loaves and 0 cakes, with a profit of $2,000. Tell me if that is what you meant.
- "Earn": I read this as revenue, with cost subtracted. If $12 and $4 were already profits, the plan stays the same (300 loaves and 125 cakes) and the profit would be $2,700.
Concepts-only skill, so the trial gave the model only the Skill and Read tools. The model's arithmetic was checked by hand (300 loaves and 125 cakes, profit 1,625; 800 loaves if the commitment were a minimum). Nothing was compared against a no-skill run.
What it does
Teaches the model to formulate linear programs (LP), mixed-integer programs (MILP) and quadratic programs (QP) from a plain-text problem. It is concepts only: no solver API and no code.
How it works
- Classifies every sentence as parameter, constraint, decision or objective, and treats committed wording ("plans to", "operates") differently from optional wording ("may", "can choose").
- Finds implicit objectives: "determine the plan" with costs only means minimize cost; with costs and revenues it means maximize profit.
- Tells the model to ask, or to solve every plausible reading and report all outcomes, rather than silently picking one.
- Covers modeling patterns and traps: piecewise-linear profit with integer totals, cutting stock waste, goal programming, multi-period inventory, and blending with a shared mixing step.
Good for
Operations research homework and real planning problems: production, scheduling, blending, portfolio and allocation.
Limits
It formulates; it does not solve. Pairing it with a solver is up to you. NVIDIA's included report (its own evaluation, not ours) describes the QP support in cuOpt as beta.
Low risk: pure instruction files from NVIDIA's official skills repository (Apache-2.0). No scripts, no network access, no file writes, no account or key needed. The package also contains NVIDIA's own evaluation report and a signature file; we included them unchanged and did not verify the signature. Results are model-generated: check the formulation and any numbers yourself before using them for real decisions.