AlienKevin/marin-8b-instruct-sft-terminalcorpus

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Apr 13, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

AlienKevin/marin-8b-instruct-sft-terminalcorpus is an 8 billion parameter Llama 3 architecture model, fine-tuned by AlienKevin, specifically for terminal agent tasks. It leverages the marin-community/marin-8b-instruct base model and is specialized using the nvidia/Nemotron-Terminal-Corpus dataset, featuring a 32,768 token context length. This model is designed to process and generate responses within a terminal environment, aiming to improve performance on command-line and agentic operations.

Loading preview...

Model Overview

AlienKevin/marin-8b-instruct-sft-terminalcorpus is an 8 billion parameter model built on the Llama 3 architecture, fine-tuned from marin-community/marin-8b-instruct. Its primary specialization is for terminal agent tasks, achieved through supervised fine-tuning (SFT) on the nvidia/Nemotron-Terminal-Corpus dataset, which comprises 366,000 terminal agent trajectories.

Key Characteristics

  • Base Model: marin-community/marin-8b-instruct (Llama 3 8B).
  • Training Data: Fine-tuned on nvidia/Nemotron-Terminal-Corpus for 2 epochs, with a sequence length of 32,768 tokens.
  • Tokenizer: Utilizes marin-community/marin-tokenizer.

Performance and Limitations

While specifically trained for terminal interactions, the model's performance on the Terminal-Bench 2.0 benchmark shows an accuracy of 1.1% (1/89), which is significantly lower than the NemotronTerminal-8B (Qwen3-8B) reference model's 13.0% ± 2.2. This difference is attributed to architectural and tokenizer variations between Llama 3 and Qwen3. The model's TBLite progression during training reached 5% accuracy at 52% completion (Step 3000).

Usage

This model is intended for applications requiring an understanding and generation of terminal commands and agentic workflows, despite its current benchmark limitations compared to other specialized models.