ZhuofengLi/nemotron-8b-sft

Nemotron-Terminal-8B-SFT

Model Description

Qwen3-8B fine-tuned on the nemotron_terminal_filtered dataset using Supervised Fine-Tuning (SFT). This model is trained to improve terminal/coding agent capabilities, following the methodology from the paper On Data Engineering for Scaling LLM Terminal Capabilities (NVIDIA, 2026).

Final val/loss: 0.353

Training Data

• Dataset: locailabs/nemotron_terminal_filtered
• Description: Filtered terminal interaction data from the Nemotron corpus, containing high-quality coding and shell agent trajectories.
• Split: 29,850 train / 150 val (0.5% val ratio)

Training Details

Framework

• Training Framework: veRL
• Backend: FSDP1

Hyperparameters

Parameter	Value
Learning Rate	2e-5
Weight Decay	1e-4
Epochs	2
Global Batch Size	128
Micro-batch Size per GPU	1
Max Sequence Length	32,768 tokens
Optimizer	AdamW (β₁=0.9, β₂=0.95)
LR Scheduler	Cosine with 10% warmup
Gradient Clipping	1.0
Sequence Parallelism	1
Param Offload	Yes
Optimizer Offload	Yes
attn_implementation	sdpa

Hardware

Item	Value
GPU	H200
Nodes	4
GPUs per Node	8
Total GPUs	32
Approx. Training Time	~6 hours

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ZhuofengLi/nemotron-8b-sft", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("ZhuofengLi/nemotron-8b-sft")