Fix incorrect parameter count and remove duplicate cell in Lab 5.5

[Yassir-249]

Summary

Two minor edits to course_5/gdm_lab_5_5_implement_lora_for_parameter_efficient_fine_tuning.ipynb:

1. The Recap markdown describes Gemma 3 1B as having "1.3 billion parameters". The Gemma 3 Technical Report, Table 1, gives 698M non-embedding + 302M embedding = ~1.0B total for the 1B variant, which matches the model.summary() output students see in the previous lab (Total params: 999,885,952). Updating the figure to match.
2. The example-inputs code cell — the one that prints Output for ones only, Output for high 1st and 2nd input, and Output with input that reverses the sign — appears twice in a row in the "Intuition behind matrix decomposition" section. The two cells are functionally identical (only whitespace differs: x,delta_W vs x, delta_W) with no markdown between them. Removing the duplicate.

Changes

• Recap markdown: 1.3 billion parameters → 1 billion parameters.
• Cell removal: delete the second of the two identical example-inputs cells. The first remains in place, preserving reading order. delta_W is defined in the cell immediately above both copies, and the next markdown ("If you add an additional row to A...") moves to a new topic, so no downstream cell depends on the deleted one.

Verification

Tested on a Colab T4 GPU runtime against git+https://github.com/google-deepmind/ai-foundations.git@main:

• All three coding activities pass: LoraDense.call() returns shape (1, 512) matching the pretrained model; pretrained_model.summary() reports 262,656 params; lora_model.summary() reports 270,848 total / 8,704 trainable / 262,144 frozen; feedback.check_loralab_answer(0.0331, num_units=512, rank=8) returns Nice! Your answer looks correct.
• Removing the duplicate cell does not break the run — the surviving cell uses the delta_W defined immediately above it, and no later cell references a variable defined in the removed cell.

Related

• Fix T4 OOM error visibility and parameter count claim in Lab 5.4 #32 — same parameter count correction in Lab 5.4 (separate PR per the one-notebook-per-PR convention).

[Yassir-249]

Resolved on main — the duplicate cell is gone and the parameter count is corrected. Closing, thanks!