---
license: other
license_name: nvidia-open-model-license
base_model: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
tags:
- rovo
- orchestrator
- tool-calling
- agentic
- sft
- nemotron
- mamba-moe
library_name: transformers
pipeline_tag: text-generation
---

# Nemotron-3-Super-120B — Rovo Chat Orchestrator SFT (iter_001)

Full-parameter SFT of **NVIDIA Nemotron-3-Super-120B** (hybrid Mamba-2 + Latent-MoE, top-22 routing over 512 experts) for the **Rovo Chat long-horizon (LH) orchestrator** task: given a multi-turn conversation + a tool catalog, decide whether to answer in text or call the right tool with the right arguments.

## Results — 725-example LH-orchestrator answer-accuracy (Claude-Opus-4.8 judge)

| model | overall | text | tool |
|---|---|---|---|
| **This SFT (iter_001)** | **70.5%** | **78.9%** | **46.8%** |
| base Nemotron-3-Super-120B (think_off) | 68.0% | 77.2% | 42.1% |
| Qwen3-32B | 71.0% | 85.4% | 30.5% |
| Gemma-4-31B | 84.8% | 95.9% | 53.7% |

**Beats the base model on all three metrics (+2.5 overall, +1.7 text, +4.7 tool).** The largest gain is on **tool-calling — the base's weakest dimension** — and this model's tool accuracy also exceeds Qwen3-32B's.

## Training

- **Stack:** Megatron-Core 0.16.1 + Megatron-Bridge (full-parameter SFT, not LoRA).
- **Recipe:** 1 epoch (366 steps, ~2.1 h), LR 2e-6 cosine (~5% warmup), weight_decay 0.1, GBS 16, bf16; train loss 1.47→1.20 (stable).
- **Parallelism / memory:** TP=8 / EP=8 / DP=4 on 4×H200; FlashAttention 2.8.3, MoE expert `capacity_factor=1.5`, full activation recompute, CPU-offloaded distributed optimizer.
- **Data:** Rovo Chat Orchestrator SFT v3 (high-quality prod traces, CSR>90), 5,849 examples after prep.

## Known limitation (next-iteration target)

The 124.5B top-22/512-expert MoE + the ~11K-token Rovo system prompt cap full-parameter training to **seq=16384** on 4×H200, so each example was right-anchored-truncated to system + a few tools + recent context + the final assistant turn — the full ~46K-median agentic traces do not fit. The model still improved over base; remaining errors (spurious/missed/wrong tool) should shrink with the full-length context that a larger GPU budget would unlock.

*Fine-tuned on internal Atlassian Rovo production data.*