ZhuofengLi/qwen3.5-9b-nemotron-sft-ckpt300

Model Overview

This model, qwen3.5-9b-nemotron-sft-ckpt300, is an intermediate checkpoint (step 300/386) of a Qwen3.5-9B model. It has been fine-tuned by ZhuofengLi using the ms-swift framework with DeepSpeed ZeRO-3, specifically targeting improved terminal interaction and agentic capabilities. The training utilized the NVIDIA Nemotron-Terminal-Corpus dataset, which comprises 366,000 multi-step terminal execution trajectories.

Key Capabilities

• Enhanced Terminal Interaction: Optimized for understanding and generating responses within a terminal environment.
• Agentic Task Execution: Designed to handle multi-step tasks, particularly in areas like math, code, and software engineering.
• Large Context Window: Supports a maximum sequence length of 262,144 tokens during training, indicating strong contextual understanding.

Training Details

The model was trained for 2 epochs on H200 GPUs across 8 nodes (64 GPUs total) using BF16 precision. Key hyperparameters included a learning rate of 2e-5, AdamW optimizer, and a cosine LR scheduler with 10% warmup. The training dataset, nvidia/Nemotron-Terminal-Corpus, focuses on complex terminal-based problem-solving.

When to Use This Model

This model is particularly suitable for applications requiring advanced interaction with terminal environments, automating complex command-line tasks, or developing AI agents that can perform multi-step operations in software development or data analysis contexts. Its specialization in terminal trajectories makes it a strong candidate for tasks involving code execution, debugging, and system administration.