---
base_model: nvidia/Llama-3_3-Nemotron-Super-49B-v1_5
library_name: transformers
tags:
- llama-factory
- lora-merged
- bf16
- text-generation
---

# Nemotron Super 49B AIE v11 5500 Merged BF16

This is a merged bf16 export of `nvidia/Llama-3_3-Nemotron-Super-49B-v1_5` with the local AIE v11 LoRA adapter merged into the base weights.

Training summary:

- Dataset: `aie_v11`, 5,227 examples.
- Stage 1: 4-bit LoRA SFT at `cutoff_len=4096`, 2 epochs.
- Stage 2: continued from the stage 1 adapter at `cutoff_len=5500`, 2 epochs.
- Stage 2 learning rate: `3e-5`.
- Stage 2 final train loss: `0.1407`.
- LoRA target modules: `q_proj,v_proj`.
- Export dtype: `bfloat16`.

The merged model was exported with LLaMA-Factory and split into 21 safetensors shards.