README.md

metadata

license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
base_model: Qwen/Qwen2.5-Coder-7B
tags:
  - terminal-agent
  - agent
  - code

TerminalTraj-7B

This is the 7B model presented in the paper Large-Scale Terminal Agentic Trajectory Generation from Dockerized Environments.

Introduction

Training agentic models for terminal-based tasks critically depends on high-quality terminal trajectories that capture realistic long-horizon interactions across diverse domains. TerminalTraj is a scalable pipeline that:

1. Filters high-quality repositories to construct Dockerized execution environments.
2. Generates Docker-aligned task instances.
3. Synthesizes agent trajectories with executable validation code.

Using TerminalTraj, the authors curated 32K Docker images and generated 50,733 verified terminal trajectories. This model is fine-tuned from the Qwen2.5-Coder-7B backbone, achieving significant performance improvements on TerminalBench.

• Repository: multimodal-art-projection/TerminalTraj
• Paper: Large-Scale Terminal Agentic Trajectory Generation from Dockerized Environments
• Dataset: m-a-p/TerminalTraj

Sample Usage

You can use this model with the transformers library:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "m-a-p/TerminalTraj-7B"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Inference example
prompt = "Write a bash script to find all .py files in a directory and count the lines of code."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation

BibTeX:

@misc{wu2026largescaleterminalagentictrajectory,
      title={Large-Scale Terminal Agentic Trajectory Generation from Dockerized Environments}, 
      author={Siwei Wu and Yizhi Li and Yuyang Song and Wei Zhang and Yang Wang and Riza Batista-Navarro and Xian Yang and Mingjie Tang and Bryan Dai and Jian Yang and Chenghua Lin},
      year={2026},
      eprint={2602.01244},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.01244}, 
}