Exp/balanced 0.9627

[rixhavraj]

🚀 Parameter Golf Submission

Model

Balanced Peak Architecture

• Layers: 12
• Model Dim: 768
• Heads: (as configured)
• Parameters: ~7.2M
• Compressed Size: ~1.13 MB

Results

• Validation BPB: 0.9627
• Best among all experiments (EXP1, EXP2, EXP3)

Summary

This model was obtained after a 36-hour optimization cycle exploring:

• Deep Stable
• Wide Aggressive
• Balanced Peak (winner)

Balanced configuration achieved the best tradeoff between depth and width, outperforming previous runs.

Included Files

• Final model: exp_balanced_peak.ptz
• Logs: logs/exp_balanced_peak.txt
• Summary: summary.txt

Notes

• Model is fully quantized and optimized for size constraints
• No test-time tricks (pure training improvement)

---

This PR represents the best-performing model from my experiments.