Skip to content

UPDATES_SUMMARY.md

Latest commit

 

History

History
95 lines (74 loc) · 3.29 KB

UPDATES_SUMMARY.md

File metadata and controls

95 lines (74 loc) · 3.29 KB

Results Update Summary - November 19, 2025

Overview

All numbers in the LaTeX report and video presentation script have been updated to match the experimental results from the completed run in /tmp/Prascevicius_Martynas_Code/.

Actual Results (from today's run)

All Experiments Ranked by Test Accuracy:

  1. epochs_5: 90.84% (highest, but 8.1% train-test gap - overfitting)
  2. epochs_4: 90.64%
  3. lr_2e5: 90.62% ← BEST LEARNING RATE
  4. batch_16: 90.60%
  5. baseline_default: 90.59% ← BASELINE
  6. epochs_3: 90.59%
  7. lr_1e5: 90.45%
  8. lr_3e5: 90.37%
  9. batch_8: 90.30%
  10. lr_5e5: 90.21%
  11. batch_32: 90.12% ← WORST

Key Findings Changes

OLD (incorrect) vs NEW (correct):

  1. Baseline Accuracy

    • OLD: 90.77%
    • NEW: 90.59%

  2. Best Configuration

    • OLD: lr_1e5 at 91.04%
    • NEW: epochs_5 at 90.84% (but shows overfitting)
    • BEST BALANCED: lr_2e5 at 90.62% or epochs_4 at 90.64%

  3. Learning Rate Results

    • OLD: 1e-5 best (91.04%), 2e-5 worse (90.96%)
    • NEW: 2e-5 best (90.62%), 1e-5 worse (90.45%)
    • INTERPRETATION CHANGE: Validates BERT recommendations (not challenges them)

  4. Batch Size Results

    • OLD: batch_16 = 90.91%, batch_32 = 90.40%
    • NEW: batch_16 = 90.60%, batch_32 = 90.12%
    • Accuracy gap: 0.48% (similar finding)

  5. Training Duration

    • OLD: 3 epochs best (91.02%), 5 epochs worse (90.28%)
    • NEW: 5 epochs best (90.84%), 3 epochs baseline (90.59%)
    • INTERPRETATION CHANGE: Test accuracy improves despite overfitting (complex pattern)

Files Updated

1. /Users/m2000uk/Desktop/coding/AI/CW2/Prascevicius_Martynas_DistilBERT.tex

  • Abstract: Updated all key findings and numbers
  • Baseline configuration: 90.77% → 90.59%
  • Table 1: All top 5 results updated with correct metrics
  • Learning rate section: Complete reversal - now validates BERT recommendations
  • Batch size section: Updated accuracies and training times
  • Overfitting section: New interpretation about complex patterns
  • Discussion: All three key findings rewritten
  • Conclusion: Updated to reflect validation (not challenge) of BERT guidance

2. /Users/m2000uk/Desktop/coding/AI/CW2/FINAL_VIDEO_SCRIPT.md

  • Methodology slide: Baseline 90.77% → 90.59%
  • Learning rate slide: Complete reversal of findings
  • Batch size slide: Updated all numbers and times
  • Overfitting slide: New complex pattern explanation
  • Results summary: All numbers updated
  • Key findings: All three contributions rewritten
  • Slide suggestions: All updated with correct numbers

Validation

All numbers now match the JSON files in:

  • /Users/m2000uk/Desktop/tmp/Prascevicius_Martynas_Code/results/
  • /Users/m2000uk/Desktop/coding/AI/CW2/results/

These directories contain identical results (verified by diff).

Next Steps

  1. ✅ Compile LaTeX to verify no errors
  2. ✅ Check that figures match the text descriptions
  3. ✅ Review video script for consistency
  4. ✅ Ensure all percentages are rounded consistently (2 decimal places)

Student Information

  • Name: Martynas Prascevicius
  • Student ID: 001263199
  • Email: mpcode@icloud.com
  • Course: COMP1818 Artificial Intelligence Applications
  • Deadline: November 19, 2025, 5pm UK (Grace: November 21, 5pm)

Document created: November 19, 2025 All updates completed and verified