swe-world-docker-free-training

name	swe-world-docker-free-training
title	SWE-World: Building Software Engineering Agents in Docker-Free Environments
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2602.03419
keywords	["Software Engineering Agents","Training Environment","Learned Simulators","LLM-Based Surrogates","Cost Reduction"]
description	Replace Docker environments with learned LLM surrogates comprising a lightweight sandbox for deterministic file operations, a transition model predicting execution feedback, and a reward model acting as virtual test runner. Reduces training infrastructure costs while scaling SWE agent training data.

SWE-World: LLM-Based Surrogate Training Environments

Training software engineering agents typically requires expensive infrastructure to run code in isolated Docker containers. SWE-World replaces this with learned surrogates that simulate execution feedback, dramatically reducing infrastructure overhead while enabling larger-scale training data.

The key insight is that execution can be decomposed: deterministic file operations (navigation, editing) can be handled directly, while non-deterministic execution effects (stdout, exit codes, test results) are predicted by lightweight models. This hybrid approach maintains fidelity while eliminating Docker overhead.

Core Concept

SWE-World comprises three components working in tandem:

1. Lightweight Sandbox: Handles all deterministic state transitions (file system operations) directly without invoking external processes
2. Transition Model (SWT): Predicts step-level execution output—stdout, stderr, exit codes—based on file state and command
3. Reward Model (SWR): Acts as a virtual test runner, generating test reports and binary success signals

This design preserves task semantics while replacing expensive runtime dependencies with efficient learned components.

Architecture Overview

• Deterministic Sandbox: Fast, stateless file operation handling (read, write, delete, navigate) with in-memory state
• SWT (Transition Model): Small LLM fine-tuned to predict execution trace outputs given command and file context
• SWR (Reward Model): Outputs test success/failure signals, trained on real test execution traces
• Training Loop: Agents generate sequences of commands; sandbox updates file state; models predict execution effects
• Data Scaling: Larger training sets possible since models are faster than real Docker execution

Implementation

Step 1: Build the Deterministic Sandbox

Create a lightweight virtual file system that tracks state without external process invocation.

# Lightweight deterministic sandbox
class DeterministicSandbox:
    def __init__(self):
        self.file_system = {}
        self.current_dir = "/"
        .env_vars = {}

     () -> :
        
        full_path = ._resolve_path(path)
         full_path   .file_system:
             
         .file_system[full_path]

     () -> :
        
        full_path = ._resolve_path(path)
        .file_system[full_path] = content
         

     () -> :
        
        full_path = ._resolve_path(path)
        files = [f  f  .file_system.keys()
                  f.startswith(full_path)]
         .join(files)

     () -> :
        
         path.startswith():
             path
         .replace(, )