Create README.md

README.md

@@ -0,0 +1,70 @@
+---
+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:
+
+```text
+FROM ./aegisnode-validate-6.7b-unsloth.Q4_K_M.gguf
+
+TEMPLATE """{{ if .System }}{{ .System }}
+{{ end }}### Instruction:
+{{ .Prompt }}
+### Response:
+"""
+
+SYSTEM """You are a Principal AWS Cloud Architect. Write flawless, enterprise-grade Terraform code. Output ONLY valid HCL."""
+```
+
+Then 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`).