| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen2.5-14B-Instruct |
| tags: |
| - finops |
| - aws |
| - terraform |
| - cloud-governance |
| - qlora |
| - peft |
| language: |
| - en |
| - es |
| pipeline_tag: text-generation |
| library_name: peft |
| datasets: |
| - ccortezb/finoptix-training-data |
| model-index: |
| - name: FinOptix-14B |
| results: [] |
| --- |
| |
| # ☁️ FinOptix-14B |
|
|
| **Your FinOps architect in 14 billion parameters.** |
|
|
| A QLoRA fine-tuned Qwen 2.5 14B model specialized in AWS Cloud Governance, FinOps Cost Optimization, and Infrastructure as Code (Terraform HCL) auditing. |
|
|
| Developed as a submission for the **[Hugging Face Build Small Hackathon](https://huggingface.co/build-small-hackathon)**. |
|
|
| ## 🚀 Key Capabilities |
|
|
| - **Terraform HCL Auditor** — Scans IaC for cost waste, governance violations, rightsizing opportunities. Outputs compliant code. |
| - **AWS Cost JSON Parser** — Extracts anomalies, burn rates, and actionable insights from Cost Explorer / Anomaly Detection payloads. |
| - **BYaML Governance Engine** — Validates architecture YAML against a 38-type component catalog and 13 governance policies. |
| - **FinOps Q&A** — Expert answers on Reserved Instances, Savings Plans, tagging strategies, unit economics, and more. |
|
|
| ## 🛠️ Training Details |
|
|
| | Parameter | Value | |
| |-----------|-------| |
| | Base Model | | |
| | Method | QLoRA (4-bit NF4, double quantization) | |
| | LoRA Config | r=16, alpha=32, dropout=0.05, target_linear=True | |
| | Hardware | NVIDIA A100-80GB (Modal) | |
| | Duration | ~40 minutes | |
| | Epochs | 3 | |
| | Dataset | 265 gold synthetic examples (FinOps, Terraform, AWS Cost, BYaML) | |
| | Sequence Length | 2048 tokens | |
| | Optimizer | paged_adamw_8bit, lr=2e-4, cosine schedule | |
| |
| ## 📊 Dataset Composition |
| |
| | Category | Count | Description | |
| |----------|-------|-------------| |
| | Terraform HCL | 100 | Audit + refactor (rightsizing, tags, encryption, lifecycle) | |
| | AWS Cost JSON | 80 | Anomaly detection, budget alerts, cost-by-service | |
| | BYaML Governance | 50 | Schema validation, policy checks, relationship verification | |
| | FinOps Q&A | 15 | Deep expert answers (frameworks, strategies, tooling) | |
| | Bash/Python Scripts | 20 | Real boto3 + bash for cloud automation | |
| |
| All data is **100% synthetic** — no client or proprietary information. Modeled after real-world AWS patterns. |
| |
| ## 💻 Usage |
| |
| ### With PEFT (recommended) |
| |
| |
| |
| ## 🎯 Demo |
| |
| Try it live: **[FinOptix-14B Space](https://huggingface.co/spaces/build-small-hackathon/finoptix14b)** |
| |
| ## 🏕️ Hackathon Context |
| |
| - **Track:** Backyard AI — solving a real FinOps problem for my own startup (Brickstore AI) |
| - **Constraint:** ≤ 32B parameters (this model is 14B ✅) |
| - **Builder:** Carlos Cortez ([@ccortezb](https://huggingface.co/ccortezb)) — AWS Community Hero, Lima, Peru |
| - **Badges:** 🎯 Well-Tuned · 🎨 Off-Brand · 📓 Field Notes |
| |
| ## ⚖️ License |
| |
| Apache 2.0 — inheriting from Qwen 2.5 base model license. |
| |
| ## 🙏 Acknowledgments |
| |
| - [Qwen Team](https://huggingface.co/Qwen) for the exceptional base model |
| - [Modal](https://modal.com) for 50 GPU credits (training infrastructure) |
| - [Hugging Face](https://huggingface.co) for hosting and hackathon organization |
| - [Axolotl](https://github.com/axolotl-ai-cloud/axolotl) for the training framework |
| |
