Natural language to bash environment

[PLippmann]

PR Type

• RL Environment PR - Complete Environment Snapshot & Zero-Training sections

---

📝 General Information

Description

This PR adds a NL2Bash Generation Environment for training LLMs to convert natural language instructions into executable Bash commands.

Key Features:

• Uses the NL2SH-ALFA dataset (40k+ examples)
• String matching verification - Commands are verified against gold standard commands (exact or normalized match)
• Binary reward signal: 1.0 for correct matches, -1.0 for incorrect answers
• Safe for training (no execution risks) while maintaining high correlation with correctness

---

🔖 Environment Snapshot

Field	Your Entry
Environment Name	bash_env
Short Description	Train LLMs to generate Bash commands from natural language strings
Category	Verifiable-Reasoning
Dataset Needed?	Yes - westenfelder/NL2SH-ALFA (MIT License, auto-downloaded)
External Deps	None
Environmental Variables	None required
Compute Footprint Estimate	<1 GB RAM, <1s string matching verification

🧪 Zero-Training Test Results

Click to expand test results

Unit Tests: 23/23 passing

$ python -m pytest test_bash_utils.py
============================= test session starts ==============================
collected 23 items
test_bash_utils.py .......................                               [100%]
======================== 23 passed in 0.07s ========================

LLM Integration Test: Qwen2.5-1.5B-Instruct on NVIDIA A100

============================================================
Bash Environment Integration Test
============================================================
Server: http://localhost:9001/v1 (vLLM)
Model: Qwen/Qwen2.5-1.5B-Instruct
WandB: https://wandb.ai/teateam/atropos-environments_community_bash_env?nw=g3bfzo5dcc4

Examples:

✓ Good Example (Score: 1.0)

• Instruction: "Launch yazi file manager"
• Generated: yazi
• Result: Matches correct command exactly (+1.0)

✗ Bad Example (Score: -1.0)

• Instruction: "To disconnect from Tailscale..."
• Generated: tctl disconnect
• Correct: tailscale down
• Result: Mismatch (-1.0)

---

✅ Developer & Reviewer Checklist

• Code follows project style (black, isort, flake8 pass with pre-commit)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• New and existing unit tests pass locally with my changes (23/23)
• Docstrings added for all new public classes / functions
• If .env vars required, did you add it to the .env.example in repo root?

[teknium1]

Reviewing this week <3