Cloud-Gym
Scalable Training Data Generation for Infrastructure-as-Code Repair via Environment Inversion.
Cloud-Gym generates (broken_config, error_message, fix) training pairs for IaC repair by applying environment inversion — taking working Terraform, CloudFormation, and OpenTofu configs and systematically breaking them using a defined fault taxonomy. It includes a benchmark (188 entries across 8 error categories) and fine-tuned models that run entirely on CPU.
stackfix: AI-Powered IaC Repair
The stackfix CLI tool validates and repairs broken IaC files using fine-tuned local models. No API keys, no cloud costs, no data leaves your machine.
Install
pip install stackfix
Download a Model
# Recommended: 3B Q4 (1.8 GB, 87% pass@1)
python -c "
from huggingface_hub import hf_hub_download
hf_hub_download('Tetsuto/iac-repair-3b-gguf', 'iac-repair-3b-q4.gguf', local_dir='.')
"
Usage
# Check files for errors
stackfix check main.tf template.yaml
# Repair a broken file (show diff)
stackfix repair main.tf --backend gguf --model iac-repair-3b-q4.gguf
# Repair and apply fix in place
stackfix repair main.tf --apply --backend gguf --model iac-repair-3b-q4.gguf
# Explain errors in plain language
stackfix discuss main.tf --backend gguf --model iac-repair-3b-q4.gguf
# Pipe mode (stdin/stdout)
cat broken.tf | stackfix repair - --backend gguf --model iac-repair-3b-q4.gguf > fixed.tf
# Check all changed IaC files in git
stackfix git-diff --backend gguf --model iac-repair-3b-q4.gguf
Models
| Model | Size | RAM | Speed (CPU) | pass@1 | HuggingFace |
|---|---|---|---|---|---|
| 7B Q4 | 4.5 GB | ~8 GB | ~20 tok/s | 0.926 | Tetsuto/iac-repair-7b-gguf |
| 3B Q4 | 1.8 GB | ~4 GB | 49 tok/s | 0.867 | Tetsuto/iac-repair-3b-gguf |
| 0.5B Q4 | 379 MB | ~800 MB | 127 tok/s | 0.723 | Tetsuto/iac-repair-0.5b-gguf |
All models are fine-tuned Qwen2.5-Coder with LoRA, exported to GGUF. They run on any CPU (Linux, macOS, Windows).
Backends
| Backend | Install | Platform | Use Case |
|---|---|---|---|
gguf |
pip install stackfix |
Any (CPU) | Default — CI/CD, Lambda, servers |
mlx |
pip install stackfix[mlx] |
Apple Silicon | Local dev on Mac |
ollama |
pip install stackfix[ollama] + Ollama |
Any | When Ollama is already running |
CI/CD Integration
Add to your GitHub Actions workflow to catch IaC errors on every PR:
- name: Check IaC
run: |
pip install stackfix
python -c "
from huggingface_hub import hf_hub_download
hf_hub_download('Tetsuto/iac-repair-3b-gguf', 'iac-repair-3b-q4.gguf', local_dir='.')
"
stackfix check **/*.tf **/*.yaml
See examples/USE_CASES.md for more deployment scenarios (pre-commit hooks, Lambda, pipeline integration).
Pre-Commit Hook
# .pre-commit-config.yaml
repos:
- repo: local
hooks:
- id: stackfix
name: stackfix
entry: stackfix pre-commit --backend gguf --model iac-repair-3b-q4.gguf
language: python
types_or: [terraform, yaml]
additional_dependencies: ['stackfix[gguf]']
Benchmark
188 entries across 8 error categories, 3 difficulty levels, and 2 formats (Terraform + CloudFormation).
Results Summary
| Model | pass@1 | Terraform | CloudFormation | High | Medium | Low |
|---|---|---|---|---|---|---|
| 7B v2 fine-tuned | 0.926 | 0.993 | 0.750 | 0.960 | 0.897 | 0.923 |
| 3B rank4 fine-tuned | 0.867 | 0.912 | 0.750 | 0.964 | 0.797 | 0.821 |
| qwen2.5-coder:7b (base) | 0.856 | 0.905 | 0.707 | 0.840 | 0.859 | 0.893 |
| 0.5B distilled | 0.723 | 0.775 | 0.590 | 0.809 | 0.648 | 0.731 |
| llama3.2:3b (base) | 0.641 | 0.734 | 0.361 | 0.684 | 0.636 | 0.533 |
| gemma-4-26b (base) | 0.009 | 0.000 | 0.032 | 0.000 | 0.004 | 0.051 |
Fine-tuning a 0.5B model outperforms a 26B base model by 80x.
Training Data Generation
Cloud-Gym generates training data via environment inversion:
- Collect working IaC configs from GitHub, Terraform Registry, AWS samples
- Break them systematically using a fault taxonomy (28+ fault types across 8 categories)
- Validate broken configs to capture real error messages
- Pair (broken + errors) with the original working config as the gold fix
# Generate training data
cloud-gym taxonomy # View fault types
python scripts/scrape.py # Collect gold configs
cloud-gym invert # Generate broken variants
cloud-gym export # Export training pairs
Project Structure
cloudgym/
taxonomy/ Fault type definitions (28+ types, 8 categories)
scraper/ Gold config collection
validator/ IaC validation wrappers (terraform, cfn-lint)
inverter/ Fault injection engines
generator/ Training data pipeline
benchmark/ Evaluation harness
fixer/ stackfix CLI tool + model backends
scripts/ Training, evaluation, and export scripts
examples/ Broken IaC examples + use case docs
Supported Formats
- Terraform (
.tf) — validated withterraform validate - CloudFormation (
.yaml,.yml,.json) — validated withcfn-lint - OpenTofu (
.tf) — same as Terraform
License
MIT
Metadata
Release files for stackfix 0.1.2
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| stackfix-0.1.2.tar.gz | 483.5 kB | Without provenance | Details for stackfix-0.1.2.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| stackfix-0.1.2-py3-none-any.whl | Python 3 | none | any | Details for stackfix-0.1.2-py3-none-any.whl |
Total release size: 549.4 kB
Release files / stackfix-0.1.2.tar.gz
| Download URL | stackfix-0.1.2.tar.gz |
|---|---|
| Size | 483.5 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
eb67f692849b0a6a65ebe64fd603cd368f6359680c302d9ec7792c6737740c88
|
|
BLAKE2b-256 checksum How to use checksums |
db97cddb6a7b4f9aed6127d884817a179e247902b195a235c887c834c3e49f95
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.13
|
Release files / stackfix-0.1.2-py3-none-any.whl
| Download URL | stackfix-0.1.2-py3-none-any.whl |
|---|---|
| Size | 65.8 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
37fc5d16ae3e20394d8db3de449bf88dd90c4dcdb4f18c23734ba8b85b061b11
|
|
BLAKE2b-256 checksum How to use checksums |
63206e6d6a56d00fb543bddc4770496c0c716c91f70243841b5f02171aabaa00
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.13
|
Security for stackfix 0.1.2
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.