Nemotron-3-Super-120B — Project-Health Risk Prediction SFT (iter001b · E1)
Full-parameter SFT of NVIDIA Nemotron-3-Super-120B (hybrid Mamba-2 + Latent-MoE, 512 experts) for project-health risk prediction: given a project snapshot at a cut-off time T, the model reasons through open questions and emits a strict <PredictedRisks> block of <Risk> items (Statement / Reasoning / Tag / Severity / Time / Likelihood / Mitigation).
This is checkpoint E1 — the winner of training iteration 001b.
Results (20 held-out eval projects, greedy decode)
| Model | D1 recall | D2 precision | D3 calibration | D4 reasoning | D5 format | Total /25 |
|---|
| E1 (this model, 1 epoch) | 2.72 | 3.44 | 2.61 | 3.29 | 4.95 | 17.01 |
| base (Nemotron-3-Super-120B) | 2.60 | 3.18 | 2.36 | 3.19 | 4.78 | 16.11 |
| E2 (2 epochs) | 2.79 | 2.81 | 2.59 | 3.03 | 4.60 | 15.82 |
E1 beats the base model by +0.9/25 (wins 10/20 examples) and is degeneration-free: 0/20 repetition loops and 20/20 valid schema under greedy decoding — versus the prior SFT run, which looped catastrophically (one output repeated a phrase ×159 to the 32K cap). Scored per an internal risk-prediction rubric (D1–D5).
Training
- Recipe: 1 epoch, LR 2e-6 (cosine, ~5% warmup), weight_decay 0.1, seq_length 10240, bf16.
- Stack: Megatron-Core + Megatron-Bridge, TP=8 / EP=8 / DP=2 on 16×H200, CPU-offloaded distributed optimizer.
- Key fix vs the prior run: corrected the chat-SFT loss mask so loss is computed on assistant tokens only (the prior run trained on the full prompt, which — together with over-training — caused greedy repetition/non-termination). 1 epoch is the sweet spot; 2 epochs over-trains (worse precision + a degeneration).
Intended use & limitations
- Use: project-health risk forecasting / triage assistance from structured project snapshots.
- Known limitation (next-iteration target): like the base model, it tends to over-rate severity (CRITICAL/HIGH) on projects that ultimately resolve benignly, and under-recalls long-tail/low-severity risks.
- Inference: use the model's chat template with thinking enabled; greedy decoding is safe.
Fine-tuned on internal project data — private.
