| --- |
| tags: |
| - terraform |
| - aws |
| - infrastructure-as-code |
| - hcl |
| - unsloth |
| - deepseek-coder |
| - lora |
| license: apache-2.0 |
| language: |
| - en |
| base_model: |
| - deepseek-ai/deepseek-coder-6.7b-instruct |
| --- |
| |
| # 🛡️ AegisNode Validate (6.7B) |
|
|
| AegisNode Validate is a specialized code-generation model fine-tuned to write syntactically flawless, zero-yapping AWS Terraform (HCL). It is built on top of `deepseek-coder-6.7b-instruct` using Unsloth and QLoRA. |
|
|
| This model is **Phase 1** of a larger Curriculum Learning pipeline. It has been strictly trained to master the "grammar" of Terraform, complex referencing (`depends_on`, `lifecycle`), and strict adherence to the AWS Provider ~> 5.0 format. |
|
|
| ## 🚨 CRITICAL WARNING: SYNTAX ONLY 🚨 |
| **This model has ONLY been trained against `terraform validate`.** |
| While the output will be structurally and syntactically perfect HCL, **it is not guaranteed to pass `terraform plan` or deploy successfully.** * It may hallucinate AWS region constraints (e.g., placing CloudFront WAFs outside `us-east-1`). |
| * It may create logically orphaned resources (e.g., generating a KMS key but forgetting to attach it to a database). |
| * It has **not** yet been trained on Checkov/tfsec security policies. |
|
|
| **Do not deploy this code to production without human review.** This model is intended to be used as a high-speed bootstrapping tool or a "Teacher Model" for generating training data for more advanced logic pipelines. |
|
|
| ## 🧠 Model Behavior: The "Zero-Yapping" Guarantee |
| Unlike standard conversational LLMs, AegisNode Validate has been trained on a heavily filtered dataset to completely eliminate conversational filler. |
| * It will not say "Here is your code." |
| * It will not apologize. |
| * It will not output markdown wrappers (````hcl````) unless explicitly prompted. |
| * **It outputs RAW, executable HCL from the very first token.** |
|
|
| ## 💻 Usage (Ollama / GGUF) |
|
|
| Because this model relies on the native DeepSeek-Coder template, you must use the correct instruction formatting. If you download the `.gguf` file, use the following `Modelfile` to run it in Ollama: |
|
|
| Create and run the model: |
| ```bash |
| ollama create aegisnode-validate -f Modelfile |
| ollama run aegisnode-validate "Create a VPC in us-east-1 with CIDR 10.0.0.0/16 and two public subnets." |
| ``` |
|
|
| ## 📊 Training Details |
| * **Base Model:** `deepseek-ai/deepseek-coder-6.7b-instruct` |
| * **Dataset:** 3,470 meticulously refined and augmented Terraform trajectories. |
| * **Hardware:** 1x NVIDIA RTX 5070TI (32GB VRAM) |
| * **Framework:** [Unsloth](https://github.com/unslothai/unsloth) + Huggingface TRL |
| * **Hyperparameters:** Rank 8, Alpha 16, LR 2e-5, Cosine Decay, 1 Epoch. (Trained explicitly on Assistant responses only). |
|
|
| ## 🚀 The AegisNode Roadmap |
| - [x] **Phase 1 (AegisNode Validate):** Master HCL syntax and formatting (`terraform validate`). |
| - [ ] **Phase 2 (AegisNode Plan):** Master AWS API logic and state relationships (`terraform plan`). |
| - [ ] **Phase 3 (AegisNode Hiraya):** Master enterprise security and compliance (`checkov` / `tfsec`). |
