Releases: arnavd371/Project-Kaan
Release list
v3.1.1 — Comment cleanup
Strip decorative section banners and narrating comments from Python sources. Keep leakage warnings, Balingbing citations, and librosa parity notes. Web meta description no longer says “AI-Powered”.
Sole author: Arnav Dhiman.
v3.1.0 — Workshop-competitive research release
Summary
Workshop-facing research release on top of the v3 distilled production model.
- Multi-seed advanced suite (42/43/44) with bootstrap 95% CIs — robustness, calibration, hierarchical heads, SSL
- Hierarchical fine-tune (strict gate): 97.15% vs 96.84% baseline (seed 42)
- CCAI @ NeurIPS 2026 Papers-track draft in
workshop/ LIMITATIONS.md— IRRI ≠ phone-on-bag; soft 84.51% reference- Apache-2.0 — filled LICENSE appendix + detailed NOTICE;
web/aligned from MIT
Copyright © 2026 Arnav Dhiman (arnavd371@gmail.com). Sole GitHub contributor.
Key numbers
| Metric | Value |
|---|---|
| Baseline acc (3-seed mean) | 96.73% ± 0.37% |
| SSL acc (3-seed mean) | 96.62% ± 0.18% |
| Phone-band robustness | 60.9% ± 8.4% |
| SNR ≤10 / hard combos | ~7.6% |
| Hier fine-tune (seed 42) | 97.15% |
Paths
- Aggregate:
experiments/results/advanced_multiseed/ - Hier FT:
experiments/results/hier_finetune/ - Paper:
workshop/kaan_ccai_neurips2026.pdf - Changelog:
CHANGELOG.md
v3.0.0 — Distilled production model
Summary
- Distilled production CNN from
gbdt+extratrees+cnn_deepsoft labels (Kaggle T4) - Seed-42 val: 97.15% (macro F1 0.977) vs hard-only 95.57% and teacher ensemble 95.89%
- Shipped INT8 TFLite (~333 KB) + ONNX for
web/ - Apache
LICENSE/NOTICEfilled with Arnav Dhiman contact details - Distill pipeline:
model/distill.py,model/export_deploy.py,experiments/kaggle/push_distill.sh
Report
See experiments/results/distill/distill_report.md.
Not in this release
Field microphone corpus / phone-on-bag calibration (still desk-bound IRRI + ambient clean).
Kaan v2.0.0
Summary
Multi-approach bake-off (11 models), multi-seed stats, audits, and findings.
Production phone app path unchanged (INT8 / ONNX mel-CNN).
Approaches
cnn_shallow, cnn_deep, cnn1d, yamnet_probe, svm_rbf, mlp, gbdt, rf, extratrees, knn, logreg
Multi-seed accuracy (seeds 42/43/44 vs 84.51% ref)
| Approach | Acc mean ± std | Seeds > ref |
|---|---|---|
| gbdt | 95.36% ± 1.50 | 3/3 |
| cnn_deep | 95.15% ± 0.48 | 3/3 |
| extratrees | 94.94% ± 1.10 | 3/3 |
| svm_rbf | 94.73% ± 1.20 | 3/3 |
| logreg | 94.09% ± 1.59 | 3/3 |
| rf | 93.67% ± 0.95 | 3/3 |
| cnn_shallow | 93.57% ± 1.28 | 3/3 |
| mlp | 92.09% ± 0.32 | 3/3 |
| knn | 90.30% ± 0.37 | 3/3 |
| yamnet_probe | 85.65% ± 1.83 | 2/3 |
| cnn1d | 64.77% ± 20.7 | 1/3 |
Assets
stats.md/stats.json/aggregate_metrics.json/per_seed_metrics.jsonfindings_seed{42,43,44}.mdMODELS.md
Author: Arnav Dhiman. Apache-2.0.
CNN ablation summary table
Ablation comparison table. Author: Arnav Dhiman.
Ablation: baseline CNN
Ablation: baseline CNN
Author: Arnav Dhiman.
Ablation: without label smoothing
Ablation: without label smoothing
Author: Arnav Dhiman.
Ablation: without class weights
Ablation: without class weights
Author: Arnav Dhiman.
Ablation: without SpecAugment
Ablation: without SpecAugment
Author: Arnav Dhiman.
Ablation: full CNN recipe
Ablation: full CNN recipe
Author: Arnav Dhiman.