I85PREP
⌘ K
NUST NET study materials
NUST NET study resources

NUST NET practice questions

One question at a time. Build your NUST NET understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Engineering

A simply supported beam of span 6 m carries a central point load of 20 kN. Ignoring self-weight, what is the maximum bending moment in the beam?

  1. 30 kN·m Correct answer.
  2. 120 kN·m
  3. 20 kN·m
  4. 60 kN·m

Explanation

For a simply supported beam with a central point load P over span L, the maximum moment occurs at midspan and equals PL/4. Substituting P = 20 kN and L = 6 m gives 20 × 6 / 4 = 30 kN·m. The reaction at each support is 10 kN, and the moment at midspan is 10 kN × 3 m = 30 kN·m, which confirms the formula.

Question 2Engineering

In a steady two-dimensional incompressible flow, the continuity equation in Cartesian coordinates requires which relationship between velocity components u and v?

  1. ∂u/∂y + ∂v/∂x = 0
  2. ∂u/∂x − ∂v/∂y = 0
  3. ∂u/∂x + ∂v/∂y = constant
  4. ∂u/∂x + ∂v/∂y = 0 Correct answer.

Explanation

Conservation of mass for an incompressible fluid reduces to the divergence of the velocity field being zero. In two dimensions this divergence is ∂u/∂x + ∂v/∂y, so it must vanish everywhere in the flow field. The condition is a local statement valid at every point, not a global constant.

Question 3Engineering

Which crystal structure has the highest atomic packing factor among common metallic structures?

  1. Face-centred cubic (FCC) Correct answer.
  2. Diamond cubic
  3. Simple cubic
  4. Body-centred cubic (BCC)

Explanation

The atomic packing factor expresses the fraction of unit-cell volume occupied by atoms. FCC and HCP both reach about 0.74, the densest possible arrangement of equal spheres, while BCC reaches about 0.68 and simple cubic only about 0.52. Among the options listed, only FCC attains the maximum value of 0.74.

Question 4Engineering

A resistor of 220 Ω dissipates 5 W. What is the approximate current through it?

  1. 0.023 A
  2. 1.1 A
  3. 22.7 A
  4. 0.15 A Correct answer.

Explanation

Power dissipated in a resistor is P = I²R, so the current is I = √(P/R). Substituting gives I = √(5/220) ≈ √0.0227 ≈ 0.151 A. Checking with P = I²R: (0.151)² × 220 ≈ 5 W, which confirms the result.

Question 5Engineering

In a first-order RC circuit with zero initial capacitor voltage, the capacitor voltage after one time constant following application of a DC source of V volts is closest to which value?

  1. 0.37 V
  2. 0.63 V Correct answer.
  3. 1.00 V
  4. 0.50 V

Explanation

The charging response is v(t) = V(1 − e^(−t/RC)). At t equal to one time constant RC, the factor becomes 1 − e^(−1) ≈ 1 − 0.368 = 0.632, so the capacitor reaches about 63 percent of the source voltage. The time constant is the time required to reach this fraction, regardless of the source value.

Question 6Engineering

Which thermodynamic cycle is the ideal reference cycle for a spark-ignition reciprocating engine?

  1. Diesel cycle
  2. Brayton cycle
  3. Otto cycle Correct answer.
  4. Rankine cycle

Explanation

The Otto cycle models heat addition at constant volume, which approximates combustion in a spark-ignition engine after the fuel-air mixture is compressed. Its thermal efficiency depends on the compression ratio as 1 − r^(1−γ). The Diesel cycle uses constant-pressure heat addition and models compression-ignition engines instead.

Question 7Engineering

The SI unit of the universal gravitational constant G is best expressed as which combination of base units?

  1. m s⁻²
  2. m² kg⁻¹ s⁻²
  3. N m² kg⁻¹
  4. m³ kg⁻¹ s⁻² Correct answer.

Explanation

From Newton's law of gravitation, F = G m₁m₂/r², we get G = F r²/(m₁m₂). Substituting units gives (kg m s⁻²)(m²)/kg² = m³ kg⁻¹ s⁻². This is the base-unit form; N m² kg⁻¹ is an equivalent derived-unit form but not a base-unit combination.

Question 8Engineering

A simply supported beam of rectangular cross-section has its depth doubled while the width is unchanged. By what factor does its second moment of area about the neutral axis increase?

  1. 16
  2. 4
  3. 8 Correct answer.
  4. 2

Explanation

For a rectangle of width b and depth d, the second moment of area about the centroidal axis is I = bd³/12. Because the depth is cubed, doubling d multiplies I by 2³ = 8. The bending stiffness therefore rises eightfold if the material and span stay the same.

Question 9Engineering

A steel bar of length 2 m and cross-sectional area 500 mm² carries an axial tensile load of 100 kN. Taking Young's modulus as 200 GPa, what is the elongation?

  1. 4.0 mm
  2. 0.2 mm
  3. 0.5 mm
  4. 2.0 mm Correct answer.

Explanation

Axial elongation is δ = PL/(AE). Converting units: P = 100,000 N, L = 2000 mm, A = 500 mm², E = 200,000 N/mm². Then δ = (100,000 × 2000)/(500 × 200,000) = 2.0 mm. The axial stress of 200 MPa is well within typical steel limits, so elastic behaviour is a reasonable assumption.

Question 10Engineering

Water flows through a horizontal pipe that narrows from 100 mm to 50 mm diameter. If the velocity in the wider section is 2 m/s, what is the velocity in the narrower section?

  1. 4 m/s
  2. 2 m/s
  3. 8 m/s Correct answer.
  4. 16 m/s

Explanation

Continuity for incompressible flow gives A₁V₁ = A₂V₂, so velocity scales inversely with area. The area ratio is (100/50)² = 4, so the velocity in the narrow section is 4 × 2 = 8 m/s. The horizontal orientation means elevation changes do not affect this continuity result.

Question 11Engineering

A 3-phase, 400 V, 50 Hz induction motor has a synchronous speed of 1500 rpm and runs at 1440 rpm at full load. What is the percentage slip?

  1. 10 %
  2. 6 %
  3. 4 % Correct answer.
  4. 2.5 %

Explanation

Slip is defined as s = (Ns − N)/Ns, where Ns is synchronous speed and N is rotor speed. Here (1500 − 1440)/1500 = 60/1500 = 0.04, or 4 percent. Slip rises with load and is small at rated operating conditions for a standard induction motor.

Question 12Engineering

In a balanced three-phase star-connected system with line voltage 400 V, what is the phase voltage?

  1. 200 V
  2. 400 V
  3. 693 V
  4. 231 V Correct answer.

Explanation

For a star connection, the line voltage equals √3 times the phase voltage. Dividing 400 V by √3 gives approximately 231 V, which is the standard single-phase supply value in many systems. In a delta connection the line and phase voltages would be equal instead.

Question 13Engineering

A brass component is to be joined to a steel shaft with a joint that must be dismantled for maintenance. Which joining process is most appropriate?

  1. Brazing
  2. Adhesive bonding
  3. Bolted flange connection Correct answer.
  4. Fusion welding

Explanation

A bolted connection is a mechanical fastener that can be undone without destroying the parts, which suits maintenance requirements. Welding, brazing, and structural adhesive bonding all create permanent joints. Bolting also accommodates the dissimilar metals without the thermal distortion that fusion processes would introduce.

Question 14Engineering

In a four-stroke engine, which stroke immediately follows the compression stroke?

  1. Intake stroke
  2. Power stroke Correct answer.
  3. Scavenging stroke
  4. Exhaust stroke

Explanation

The four strokes occur in the order intake, compression, power (expansion), and exhaust. After the compressed charge is ignited near top dead centre, the burning gases expand and drive the piston down in the power stroke. Exhaust follows the power stroke, and intake begins the next cycle.

Question 15Engineering

A cantilever of length L carries a point load W at its free end. What is the maximum deflection, taking EI as the flexural rigidity?

  1. WL³/(3EI) Correct answer.
  2. WL²/(2EI)
  3. WL³/(8EI)
  4. WL⁴/(8EI)

Explanation

For an end-loaded cantilever, integrating the bending moment gives the tip deflection as WL³/(3EI). The cube of the length appears because deflection depends on the second integral of the moment distribution. A simply supported beam with a central load would give WL³/(48EI) instead.

Question 16Engineering

A Carnot engine operates between a source at 600 K and a sink at 300 K. What is its thermal efficiency?

  1. 100 %
  2. 50 % Correct answer.
  3. 33 %
  4. 67 %

Explanation

Carnot efficiency depends only on the absolute temperatures of the reservoirs: η = 1 − Tc/Th. Here 1 − 300/600 = 0.5, so the efficiency is 50 percent. No real engine operating between these temperatures can exceed this value.

Question 17Engineering

A DC shunt motor draws 20 A from a 220 V supply and has an armature resistance of 0.5 Ω. What is the back EMF?

  1. 220 V
  2. 200 V
  3. 230 V
  4. 210 V Correct answer.

Explanation

In a shunt motor the field is connected in parallel, and assuming the field current is negligible here, the armature current is about 20 A. Back EMF is Eb = V − IaRa = 220 − (20 × 0.5) = 210 V. The back EMF is always slightly less than the supply voltage in motoring operation.

Question 18Engineering

A metal matrix composite reinforced with aligned continuous ceramic fibres is loaded parallel to the fibres. Which model best predicts its longitudinal elastic modulus?

  1. Hall–Petch relationship
  2. Rule of mixtures (Voigt model) Correct answer.
  3. Gibbs–Thomson equation
  4. Inverse rule of mixtures (Reuss model)

Explanation

For loading parallel to continuous fibres, the matrix and fibres experience equal strain, so the moduli combine as a weighted average: Ec = VfEf + VmEm. This is the Voigt, or isostrain, rule of mixtures. The Reuss model applies to loading perpendicular to the fibres, where the stress is equal in both phases.

Question 19Engineering

A block of mass 5 kg rests on a horizontal surface with a coefficient of static friction of 0.4. Taking g as 10 m/s², what is the minimum horizontal force needed to start motion?

  1. 20 N Correct answer.
  2. 12.5 N
  3. 50 N
  4. 2 N

Explanation

The maximum static friction force is μsN, where the normal force on a horizontal surface equals the weight mg. Thus f = 0.4 × 5 × 10 = 20 N, which is the minimum applied force required to initiate sliding. Once motion begins, kinetic friction, which is usually lower, governs the resistance.

Question 20Engineering

A thin cylindrical pressure vessel of diameter D and wall thickness t is subjected to internal pressure p. Which statement correctly describes the stresses in the wall?

  1. Hoop stress equals longitudinal stress at pD/(4t)
  2. Longitudinal stress is pD/(2t) and is twice the hoop stress
  3. Hoop stress is pD/(2t) and is twice the longitudinal stress pD/(4t) Correct answer.
  4. Hoop stress is pD/(4t) and longitudinal stress is pD/(2t)

Explanation

For a thin-walled cylinder, equilibrium of a longitudinal section gives hoop stress σh = pD/(2t), while end-cap equilibrium gives longitudinal stress σl = pD/(4t). The hoop stress is therefore twice the longitudinal stress, which is why cylindrical vessels typically split along a longitudinal seam. The thin-wall assumption requires t to be small relative to D.

Question 21Engineering

An ideal vapour-compression refrigeration cycle uses a throttling device instead of an expansion turbine. What is the main consequence of throttling?

  1. It eliminates the need for a compressor
  2. It recovers work and raises the coefficient of performance
  3. It raises the condenser pressure above the compressor discharge pressure
  4. It is isenthalpic and dissipates available work, lowering the coefficient of performance Correct answer.

Explanation

A throttling valve is essentially isenthalpic, so the expansion occurs with no work recovery and an increase in entropy. An isentropic turbine would extract work and reduce the quality entering the evaporator, improving the coefficient of performance. Throttling is used because it is cheap and simple, despite this thermodynamic penalty.

Question 22Engineering

A control system has the open-loop transfer function G(s)H(s) = K/[s(s+2)(s+4)]. Using the Routh–Hurwitz criterion, what is the maximum value of K for closed-loop stability?

  1. 96
  2. 48 Correct answer.
  3. 24
  4. 12

Explanation

The closed-loop characteristic equation is s³ + 6s² + 8s + K = 0. For a cubic with positive coefficients, Routh–Hurwitz requires that the product of the s² and s¹ coefficients exceeds the constant term: 6 × 8 > K, so K < 48. The system is therefore stable up to K = 48, where a pair of poles crosses the imaginary axis.

Question 23Engineering

A simply supported beam is loaded by a uniformly distributed load w over its full span L. At which location does the maximum shear force occur?

  1. At the supports, with magnitude wL
  2. At midspan, with magnitude zero
  3. At midspan, with magnitude wL/8
  4. At the supports, with magnitude wL/2 Correct answer.

Explanation

For a simply supported beam with a uniformly distributed load, each support reaction equals wL/2 by symmetry. The shear force diagram falls linearly from wL/2 at one support to zero at midspan and then to −wL/2 at the other support. The maximum magnitude therefore occurs at the supports.

Question 24Computing

In a relational database, which property of a primary key guarantees that no two rows can share the same key value?

  1. Atomicity
  2. Uniqueness Correct answer.
  3. Normalization
  4. Referential integrity

Explanation

A primary key must be unique, so the DBMS rejects any insert or update that would duplicate an existing key value. Referential integrity concerns foreign-key matching, atomicity concerns transactions, and normalization concerns schema design.

Question 25Computing

A binary search is applied to a sorted array of 1,024 elements. What is the maximum number of comparisons needed to locate a target or determine it is absent?

  1. 11 Correct answer.
  2. 32
  3. 1,024
  4. 10

Explanation

Binary search halves the search space each step, so the worst case is about log2(n) rounded up. Since 2^10 = 1,024, ten halvings reduce the space to one element, and one extra comparison confirms presence or absence, giving 11.

Question 26Computing

Which number system representation is used internally by most modern CPUs to store signed integers because it makes addition and subtraction use the same hardware circuit?

  1. One's complement
  2. Sign-magnitude
  3. Binary-coded decimal
  4. Two's complement Correct answer.

Explanation

Two's complement encodes negatives so that addition works identically for signed and unsigned values, letting one adder circuit serve both. It also has a single representation of zero, unlike sign-magnitude and one's complement.

Question 27Computing

A process is described as being in the 'ready' state in an operating system. Which statement best describes that state?

  1. It is waiting for an I/O operation to complete
  2. It is currently executing on a CPU core
  3. It is able to run and is waiting for CPU allocation Correct answer.
  4. It has terminated but its entry remains in the process table

Explanation

A ready process has all resources it needs except the CPU and sits in the ready queue awaiting scheduling. Running means it currently holds a core, and waiting means it is blocked on an event such as I/O.

Question 28Computing

In the OSI reference model, at which layer does a router primarily make forwarding decisions based on logical addresses?

  1. Network layer Correct answer.
  2. Transport layer
  3. Session layer
  4. Data link layer

Explanation

The network layer handles logical addressing and routing, so routers forward packets using network-layer addresses such as IP addresses. The data link layer uses hardware addresses within a local link, while transport and session layers handle end-to-end delivery and dialogue control.

Question 29Computing

Which HTML element is semantically intended to hold self-contained content that could be distributed independently, such as a forum post or a news article?

  1. <article> Correct answer.
  2. <div>
  3. <section>
  4. <aside>

Explanation

The <article> element represents a self-contained composition that would make sense on its own, such as a blog post or forum entry. <section> groups thematically related content, <div> is a generic container, and <aside> holds tangential content.

Question 30Computing

In a spreadsheet used for computing, which reference type keeps the row fixed while allowing the column to change when a formula is copied to another cell?

  1. A$1
  2. $A$1
  3. $A1 Correct answer.
  4. A1

Explanation

The dollar sign before the column letter locks the column, while the row number without a dollar sign remains relative. Therefore $A1 keeps column A fixed but lets the row adjust as the formula is copied downward or sideways.

Question 31Computing

A stack is used to evaluate the postfix expression 6 2 3 * - . What is the result?

  1. -4
  2. 12
  3. 0 Correct answer.
  4. 18

Explanation

Reading left to right, push 6, push 2, push 3, then '*' pops 3 and 2 and pushes 6. Next '-' pops 6 and 6, computing 6 - 6 = 0, which is pushed as the final result.

Question 32Computing

Which SQL clause is evaluated to filter individual rows before any grouping is performed?

  1. GROUP BY
  2. ORDER BY
  3. HAVING
  4. WHERE Correct answer.

Explanation

WHERE filters rows before grouping, so it cannot reference aggregate functions. HAVING filters after grouping and may use aggregates, ORDER BY sorts the final result, and GROUP BY forms the groups themselves.

Question 33Computing

A function calls itself with a smaller input and has no memoization. Which consequence is most likely as the input size grows?

  1. Memory usage is always lower than an equivalent iterative solution
  2. Runtime grows exponentially for naive recursive solutions to problems like Fibonacci Correct answer.
  3. The function automatically converts to iteration at compile time
  4. The call stack depth stays constant while runtime grows linearly

Explanation

Naive recursion for problems with overlapping subproblems recomputes the same values repeatedly, giving exponential time such as O(2^n) for Fibonacci. Memoization or iteration reduces this, but without either the blow-up is expected.

Question 34Computing

What is the decimal value of the hexadecimal number 2F?

  1. 63
  2. 45
  3. 47 Correct answer.
  4. 31

Explanation

Hexadecimal 2F equals 2 times 16 plus 15, because F represents 15. That gives 32 + 15 = 47 in decimal.

Question 35Computing

Which memory in a computer system is volatile and loses its contents when power is removed but is used as the main working memory for running programs?

  1. RAM Correct answer.
  2. ROM
  3. Cache only
  4. SSD

Explanation

RAM is volatile main memory that holds programs and data currently in use and loses contents on power loss. ROM is non-volatile, SSDs are persistent storage, and cache is a smaller volatile memory that does not replace main memory.

Question 36Computing

A programmer writes a loop that reads past the end of an array in C. Which statement best explains the risk?

  1. The compiler always inserts bounds checks and throws an exception
  2. The array automatically resizes to accommodate the extra index
  3. C does not perform bounds checking, so behaviour is undefined and may corrupt memory Correct answer.
  4. The operating system guarantees the read returns zero

Explanation

C does not check array bounds at runtime, so an out-of-range index accesses memory outside the array, producing undefined behaviour that can corrupt data or crash. Languages with managed runtimes may throw exceptions, but C does not.

Question 37Computing

Which protocol is connection-oriented and guarantees reliable, ordered delivery of a byte stream between hosts?

  1. TCP Correct answer.
  2. UDP
  3. ICMP
  4. ARP

Explanation

TCP establishes a connection, uses acknowledgements and retransmission, and delivers bytes in order. UDP is connectionless and unreliable, ICMP reports network errors, and ARP resolves IP addresses to MAC addresses on a local link.

Question 38Computing

In object-oriented programming, which concept allows a subclass to provide its own implementation of a method declared in its superclass?

  1. Method overriding Correct answer.
  2. Inheritance
  3. Method overloading
  4. Encapsulation

Explanation

Overriding replaces a superclass method with a subclass version that has the same signature, enabling runtime polymorphism. Inheritance is the mechanism that makes the method available, and overloading defines multiple methods with different signatures in the same scope.

Question 39Computing

A CSS rule is written as 'p { color: blue; }' and a later rule is 'p.note { color: green; }'. A paragraph has class 'note'. Which colour applies and why?

  1. Green, because the class selector has higher specificity Correct answer.
  2. Blue, because element selectors always win
  3. Green, because class rules ignore element rules
  4. Blue, because it appears first in the stylesheet

Explanation

Specificity determines which conflicting declaration wins when selectors target the same element. A class selector combined with an element selector is more specific than a bare element selector, so green applies regardless of source order.

Question 40Computing

Which file organisation method allows records to be located by computing an address directly from a key value, giving near-constant access time?

  1. Indexed sequential organisation
  2. Sequential file organisation
  3. Hashing Correct answer.
  4. Linked-list organisation

Explanation

Hashing applies a hash function to the key to compute a storage address, so average lookup time is close to constant. Sequential files require scanning, indexed sequential files use an index plus sequential search, and linked lists require traversal.

Question 41Computing

A recursive function with a base case is called with a large input and the program terminates with a stack overflow. Which is the most likely cause?

  1. The compiler cannot optimise the base case
  2. The recursion depth exceeds the call stack limit because each call adds a frame Correct answer.
  3. The base case is reached too early
  4. The function returns too many values at once

Explanation

Each recursive call pushes a stack frame, and deep recursion can exceed the stack's fixed size, causing overflow. The fix is usually a correct base case, reduced depth, or an iterative version.

Question 42Computing

A system uses 12-bit memory addresses. How many distinct byte-addressable memory locations can it address?

  1. 4,096 Correct answer.
  2. 2,048
  3. 1,024
  4. 12,288

Explanation

With n address bits, the number of addressable locations is 2^n. For 12 bits that is 2^12 = 4,096 distinct addresses.

Question 43Computing

Which scheduling algorithm can cause starvation of long jobs because it always selects the process with the smallest estimated remaining time?

  1. Round Robin
  2. Shortest Remaining Time First Correct answer.
  3. First Fit
  4. First Come First Served

Explanation

Shortest Remaining Time First always picks the job with the least remaining time, so long jobs can be repeatedly overtaken and starve. FCFS is order-based, Round Robin gives every process a time slice, and First Fit is a memory-allocation strategy.

Question 44Computing

A relation is in third normal form (3NF). Which condition must hold in addition to being in second normal form?

  1. Every non-prime attribute depends transitively on the primary key
  2. Every attribute is part of a composite key
  3. No non-prime attribute is transitively dependent on the primary key Correct answer.
  4. All foreign keys are removed

Explanation

3NF removes transitive dependencies, so every non-prime attribute must depend directly on the key rather than through another non-prime attribute. Being in 2NF already removes partial dependencies on part of a composite key.

Question 45Computing

Two processes each hold one resource and both request the resource held by the other. Which condition of deadlock does this scenario directly illustrate?

  1. Mutual exclusion
  2. Preemption
  3. Circular wait Correct answer.
  4. Starvation

Explanation

A circular wait exists when each process holds a resource the next process needs, forming a cycle. Mutual exclusion is about exclusive resource use, and deadlock requires no preemption, not the presence of preemption.

Question 46Computing

A hash table uses chaining and stores n keys in m buckets. What is the expected average time to search for a key when the hash function distributes keys uniformly and the load factor is small?

  1. O(n)
  2. O(log n)
  3. O(n log n)
  4. O(1) Correct answer.

Explanation

With uniform hashing and a small load factor, each bucket holds few entries on average, so the expected search cost is constant. The worst case degrades to O(n) if all keys collide, but the question asks for the expected case.

Question 47BS Robotics & Artificial Intelligence

In a robotics context, which quantity does a potentiometer used as a joint sensor directly measure?

  1. Torque applied at the joint
  2. Magnetic flux density around the motor
  3. Angular position of the joint Correct answer.
  4. Angular velocity of the joint

Explanation

A potentiometer is a resistive divider whose wiper voltage changes with shaft rotation, so its output maps directly to angular position. Velocity and torque require differentiation or separate sensors, and flux density is unrelated to the resistive element.

Question 48BS Robotics & Artificial Intelligence

A robot arm's end-effector pose is described by a 4x4 homogeneous transformation matrix. Which information does this matrix contain?

  1. Neither rotation nor translation, only joint angles
  2. Both rotation and translation of the end-effector Correct answer.
  3. Only the translation of the end-effector
  4. Only the rotation of the end-effector

Explanation

A homogeneous transformation packs a 3x3 rotation block and a 3x1 translation vector into one 4x4 matrix with a final row of 0 0 0 1. This lets a single matrix multiplication both rotate and translate a point between frames.

Question 49BS Robotics & Artificial Intelligence

Which statement best describes supervised learning?

  1. The model learns patterns from labelled input-output pairs Correct answer.
  2. The model memorises the test set before training
  3. The model learns by trial and error using reward signals
  4. The model clusters unlabelled data into groups

Explanation

Supervised learning trains a mapping from inputs to known target outputs using labelled examples. Clustering is unsupervised, reward-driven trial and error is reinforcement learning, and memorising test data would be data leakage rather than a learning paradigm.

Question 50BS Robotics & Artificial Intelligence

In a feedforward artificial neural network, what is the role of an activation function such as ReLU?

  1. It stores the weights between successive layers
  2. It normalises the input features before training
  3. It shuffles the training data each epoch
  4. It introduces non-linearity so the network can model complex mappings Correct answer.

Explanation

Without a non-linear activation, stacking linear layers collapses into a single linear transformation. ReLU and similar functions bend the output so the network can approximate non-linear functions. Weights store parameters and data shuffling is a training-loop detail.