Practice: Explain And Check A Model Call

Use the worksheet to test which quantities and claims the model mental model can support. We will calculate representations and prediction loss, budget weights and context, then inspect a controlled policy answer. Each question has its own pause before the explanation.

Keep the supplied numbers, units and assumptions beside your working. These are classroom exercises, not observations from model calls, and you can complete them without buying or running a service. Start with a vocabulary containing 30,000 entries and an embedding table with 512 coordinates in each row.

How many scalar entries are in this simple matrix? State the multiplication, then explain whether this gives the total parameter count of a complete model. Pause and record your answer before continuing. The table contains 30,000 times 512, or 15,360,000 scalar entries.

It is 1 matrix. A full model can include additional projection, feedforward and other parameter collections, so this count is not its total. Keep entries separate from bytes: a storage calculation would additionally require the representation used for each entry and the relevant overhead. Now use 3 allowed source

positions with equal attention scores. Their values are (3,0), (0,3) and (0,0). Calculate the normalized shares and the combined 2-coordinate output. Show which values contribute to each coordinate. Pause and finish the calculation before opening the key.

Equal scores produce equal normalized shares, each 1 divided by 3. The 1st coordinate is 3 divided by 3, and the 2nd is also 3 divided by 3, giving (1,1). All 3 positions have a share, although the 0 vector adds no coordinate value. This is a source-position value mixture,

not the probability that an output answer is true. For the next exercise, a fixed observed target token's probability rises from 0.1 to 0.8. Using negative natural log probability, does the loss rise or fall? Calculate both values to 3

decimal places and state what the comparison says about that target. Pause and complete your working. The loss falls from approximately 2.303 to 0.223. Assigning more probability to the same observed target receives a smaller penalty under this objective.

It does not independently verify a new factual claim or supply a whole training update. The gradients, optimizer and other training conditions remain outside these 2 supplied probability cases. A checkpoint has 10 billion parameters stored at 2 bytes per parameter. Calculate its raw weight

storage in decimal gigabytes. Would a device with exactly that memory capacity necessarily hold the whole running service? Pause and record both the number and what it excludes. The multiplication gives 20 billion bytes, or 20 gigabytes in decimal units. That is raw weight

storage under the supplied precision. Runtime cache, activations, buffers and overhead require additional accounting; training adds its own state. The weight-only calculation therefore does not establish that the running service fits into 20 gigabytes of device memory. A hypothetical service has a combined 12,000-token input-and-output limit.

Reserve 2,000 tokens for output and 1,500 for the defined other input. How much remains for documents and history before extra formatting or margin? Pause and write the subtraction, its units and its assumptions. The remainder is 12,000 minus 2,000 minus 1,500, or 8,500 tokens.

This answer uses the stipulated combined accounting convention. Extra message formatting or margin has not yet been allocated. A real service may use a different convention, and fitting the packet still does not guarantee that every supplied detail is used correctly.

An MoE headline reports 60 billion total parameters and 12 billion active parameters. Can the full weight-storage plan use only the active count? Explain how the remaining weights and the serving architecture affect your answer. Pause and complete that explanation before continuing. The active count describes parameters used

under the model's per-token reporting convention. Other relevant weights still need to be accessible through the actual deployment arrangement. They may be distributed or moved under a designed system, which introduces resource and communication considerations. Plan total weights and runtime state from the configuration;

the active number alone is insufficient. Finish with a fictional policy task. Standard products can be returned within 30 days, while customized products are excluded unless defective. The complete case is customized, nondefective and delivered 10 days ago. A receives only the short 10-day question;

B receives the full policy and facts, with the same model weights. B's proposed object has eligible set to the boolean true. In a separate supplied next-token sketch, the supplied scores for true and false are 2 and 1, so greedy selection favors true.

Identify A's expected behavior, B's evidence-supported decision, whether the boolean type is valid, and what this comparison can establish. Pause and complete all parts. A should obtain the missing policy and product facts or qualify its response. B's supported decision is no, so eligible should be false

with the customization exclusion explained. The proposed true is a valid boolean type but the wrong decision under this case. A favored token can still be wrong for the task; greedy selection is not a fact checker. This A/B comparison changes input completeness,

not the underlying weights. Observed performance remains unknown because no model was run. Review your worksheet by tracing each conclusion back to its inputs and rule. Matrix entries, value coordinates, loss, bytes, context tokens and active computation have different meanings. The final policy exercise connects

model generation to evidence and validation. Carry this discipline into the independent project: define cases and expected behavior first, keep observed results blank until measured, and make only the claims your controlled evidence supports.