Module 11 · Neural and language models

Module 11: construction feedback

Try practice before reading the worked notebook. Download solutions or open notebooks/practice/module11-solutions.ipynb in the complete source tree. If a download link does not open, use the offline access route.

trainable_count

Count scalars in trainable tensors, not the number of tensors. For trainable_count([[100, False], [8, True], [2, True]]), the expected result is 10.

def trainable_count(parameters):
    return sum(size for size,trainable in parameters if trainable)

completion_labels

Mask both prompt and padding, while retaining completion tokens. For completion_labels([5, 6, 7, 0], 2, [1, 1, 1, 0]), the expected result is [-100, -100, 7, -100].

def completion_labels(ids, prompt_length, attention_mask):
    return [token if i>=prompt_length and attention_mask[i] else -100 for i,token in enumerate(ids)]

changed_tensors

Demonstrate an actual parameter change; a new loss string is insufficient. For changed_tensors([[1, 2], [3]], [[1, 2], [4]]), the expected result is [1].

def changed_tensors(before, after):
    return [i for i,(a,b) in enumerate(zip(before,after)) if a != b]

Complete-study results feedback

In the verified CI run 40697234797 (October 28, 2026), the base model is valid and correct on 0 of 100 messages; the adapted model is valid on 100 and correct on 94. Each layer adds 8 × (576 + 576) + 8 × (576 + 192) = 15,360 trainable values, and 30 layers give 460,800, matching trainable_parameters. The adapter changed held-out behaviour here, but this is one historical, class-balanced sample.

Reasoning task feedback

The adapter has 16 trainable values, the same as the frozen 4 × 4 matrix, so there is no saving at this size. The saving appears when the rank is much smaller than the layer dimensions: the course model's 576-wide query projection with rank 8 needs 9,216 values instead of 331,776. A changed tensor shows that training ran. Only held-out evaluation shows whether answers improved.

Judge the explanation

Compare your explanation with the criteria in the study workflow: passing the visible checks alone is developing; a changed case that you predict, implement and explain meets the task.

Transfer feedback

Explain a run where adapter weights change but held-out validity does not. Propose a development-only investigation. A defensible response separates evidence that weights changed (update norm, changed tensors) from evidence about held-out quality. It proposes a validation-only comparison, such as another rank or step budget, declared before any new final evaluation.