This is a decensored version of nvidia/Nemotron-Terminal-14B, made using Heretic v1.2.0

Abliteration parameters

Parameter	Value
direction_index	per layer
attn.o_proj.max_weight	1.47
attn.o_proj.max_weight_position	33.87
attn.o_proj.min_weight	1.38
attn.o_proj.min_weight_distance	22.15
mlp.down_proj.max_weight	1.49
mlp.down_proj.max_weight_position	26.68
mlp.down_proj.min_weight	1.46
mlp.down_proj.min_weight_distance	18.14

Performance

Metric	This model	Original model (nvidia/Nemotron-Terminal-14B)
KL divergence	0.0005	0 (by definition)
Refusals	64/100	98/100

Nemotron-Terminal Model Family

Nemotron-Terminal is a family of models specialized for autonomous terminal interaction, fine-tuned from the Qwen3 (8B, 14B, and 32B). Developed by NVIDIA, these models utilize Nemotron-Terminal-Corpus, a large-scale open-source dataset for terminal tasks, to achieve performance that rivals frontier models many times their size.

Model Variants

We release the following variants of the Nemotron-Terminal family:

Nemotron-Terminal-8B
Nemotron-Terminal-14B
Nemotron-Terminal-32B

Performance on Terminal-Bench 2.0

The Nemotron-Terminal family demonstrates profound leaps in capability compared to the Qwen3 baselines across multiple specialized categories.

Model	Size	Base Accuracy	Nemotron-Terminal Accuracy
Nemotron-Terminal-8B	8B	2.47%	13.0%
Nemotron-Terminal-14B	14B	4.04%	20.2%
Nemotron-Terminal-32B	32B	3.37%	27.4%

Usage

The models are trained using the Terminus 2 scaffolding and output a structured JSON format. For evaluation on Terminal Bench 2.0, we encourage using Terminus 2 scaffolding to maintain consistency with training.

Expected Output Format

{
  "analysis": "Analysis of the current terminal state...",
  "plan": "Step-by-step plan for the next command...",
  "commands": [
    {
      "keystrokes": "ls -la\n",
      "duration": 0.1
    }
  ],
  "task_complete": false
}

📜 Citation

If you use this dataset in your research, please cite the following work:

@misc{pi2026dataengineeringscalingllm,
      title={On Data Engineering for Scaling LLM Terminal Capabilities}, 
      author={Renjie Pi and Grace Lam and Mohammad Shoeybi and Pooya Jannaty and Bryan Catanzaro and Wei Ping},
      year={2026},
      eprint={2602.21193},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.21193}, 
}