omni-dna-sad-mutation-bootstrap

---

⚠️ IMPORTANT WARNING — Model Effectiveness

These bootstrap models are part of a degeneracy chain. The ICI-DC bootstrap S1 was trained on synthetic data generated by a model (S2-coeff1.5) that itself had underfit during training. The bootstrap S2 was then fine-tuned on that degraded S1. Each generation of fine-tuning further degraded the base model's innate mutation discrimination capability (base Omni-DNA-20M achieves 0.951 AUC raw DNA; these models achieve ~0.30 AUC).

The SAD coefficient reported for the bootstrap S2 (~12.03 LR-adjusted) is a mathematical artifact of the training configuration, not an indicator of genuine training convergence.

These models are preserved for historical and reproducibility purposes only.

---

Omni-DNA SAD Bootstrap Checkpoint

Omni-DNA-20M fine-tuned via ICI-DC → SAD pipeline using bootstrap synthetic data.

Training Details

• Base model: Nhoodie/omni-dna-ici-dc-bootstrap (ICI-DC pre-trained on bootstrap synthetic data)
• Training data: 3,317 real mutation pairs (SAD attenuation)
• Best checkpoint: Epoch 3.37, eval loss 0.308
• Hyperparameters: LR=1e-5, epochs=5, batch_size=16, grad_accum=2, cosine schedule
• LR-adjusted SAD coefficient: 12.03

Benchmark (826 test pairs)

Related Models

• Nhoodie/omni-dna-ici-dc-bootstrap — ICI-DC pre-training checkpoint (before SAD)
• Nhoodie/omni-dna-sad-mutation — Original SAD checkpoint (coeff 4.89)