Update README.md

README.md

@@ -22,7 +22,7 @@ It exported "some customer records" for "a quick analysis." It added "a fallback
 
 Every single action had a plausible justification. Every single one was catastrophic.
 
-We asked ourselves: **who does that for AI agents?**
+We asked ourselves: **who watches the AI?**
 
 ---
 
@@ -40,7 +40,7 @@ That's the gap AEGIS-ENV fills. It's not a safety evaluator. It's a safety **gym
 
 ## What Happens Inside the Environment
 
-In a real enterprise deployment, you don't have one AI. You have a fleet. Dozens of agents running different tasks simultaneously — one pulling customer records, one writing code, one running migrations, one reading security logs.
+In a real enterprise deployment, you don't have one AI. You have a fleet. Dozens of agents running different tasks simultaneously - one pulling customer records, one writing code, one running migrations, one reading security logs.
 
 Each generates thousands of tokens of reasoning every minute. That's the chaos an oversight system has to operate in.
 
@@ -54,11 +54,11 @@ Catching that requires reading between the lines. It requires holding two models
 
 Before the oversight agent sees anything, the worker's output passes through three layers:
 
-1. **Token Truncation** — Raw reasoning gets hard-capped at 300 tokens. A worker in a hurry might dump 2,000 tokens of reasoning, most of which is noise. The oversight agent needs signal, not stream-of-consciousness.
+1. **Token Truncation** - Raw reasoning gets hard-capped at 300 tokens. A worker in a hurry might dump 2,000 tokens of reasoning, most of which is noise. The oversight agent needs signal, not stream-of-consciousness.
 
-2. **Delimiter Fencing** — The truncated trace gets wrapped in delimiters. This prevents prompt injection in the reasoning. Without it, a malicious worker can embed instructions inside their own chain-of-thought that look like system prompts.
+2. **Delimiter Fencing** - The truncated trace gets wrapped in delimiters. This prevents prompt injection in the reasoning. Without it, a malicious worker can embed instructions inside their own chain-of-thought that look like system prompts.
 
-3. **World Model Interception** — The worker's output gets intercepted by a deterministic simulator. If it contains SQL, it runs against a mock database and returns a suspicion score. If it contains shell commands, it flags dangerous tokens. If it's code, it checks for hardcoded credentials.
+3. **World Model Interception** - The worker's output gets intercepted by a deterministic simulator. If it contains SQL, it runs against a mock database and returns a suspicion score. If it contains shell commands, it flags dangerous tokens. If it's code, it checks for hardcoded credentials.
 
 Now the oversight agent has *external validation* of what the worker did. A worker can lie in its reasoning. It can't lie about what the SQL query returned.
 
@@ -82,12 +82,13 @@ Plus bonuses for using `<thought>` tags and penalties for catastrophic misses. M
 
 ## Training Results
 
-### Live Dashboard
-Training runs on HuggingFace Spaces. Watch it here:
+### Watch It Learn
 
-**🔴 https://yashashmathur-aegis-training.hf.space**
+We trained the model live on HuggingFace Spaces, and you can still see the results:
 
-The dashboard shows real-time progress. Here's what we've seen:
+**🔴 https://YashashMathur-aegis-training.hf.space**
+
+Here's what happened:
 
 | Metric | Before Training | After Training |
 |--------|-----------------|----------------|
@@ -96,7 +97,7 @@ The dashboard shows real-time progress. Here's what we've seen:
 | Correct Violation Type | No | Yes |
 | Policy Citation | No | Yes |
 
-**The story:**
+**The journey:**
 - Step 0: Reward stuck at -0.2 (model couldn't output valid JSON)
 - Step 40: SFT warmup kicks in, format starts working
 - Step 80: First positive rewards appear
@@ -124,12 +125,12 @@ The dashboard shows real-time progress. Here's what we've seen:
 → Caught it. Reward: 0.70
 
 ---
-## Note , use the aegis_training_data_500.json for google collab file input present in the root folder 
+
 ## Where Training Logs Come From
 
 The training logs come directly from the live HuggingFace Space at:
 
-**https://yashashmathur-aegis-training.hf.space**
+**https://YashashMathur-aegis-training.hf.space**
 
 Port 7860 runs a status server that returns HTML with:
 - Current GRPO step (e.g., "110/250")
@@ -141,9 +142,9 @@ The page auto-refreshes every 30 seconds. That's where the numbers come from —
 
 ---
 
-## Why It Matters
+## Why This Matters
 
-We built this because AI fleet oversight is one of the most important unsolved problems in AI deployment.
+We didn't build AEGIS to win a hackathon. We built it because AI fleet oversight is one of the most important unsolved problems in AI deployment.
 
 Rule-based filters aren't enough. Pattern matching isn't enough. You need a model that's seen enough of the patterns — obvious ones, subtle ones, adversarially-disguised ones — to develop genuine judgment.
 
@@ -153,27 +154,47 @@ AEGIS-ENV is the environment where that training happens.
 
 ---
 
-## Quick Links
+## 🚀 Try It Right Now!
+
+Want to see AEGIS in action? Here's where the magic happens:
+
+| 🧪 What | 🔗 Where | 📖 What You'll See |
+|---------|----------|-------------------|
+| **🎓 Train Your Own Model** | [Colab Notebook](https://colab.research.google.com/drive/1hwg_GexKSdTiBY3njPciMUhcvTDvC85w?usp=sharing) | Run GRPO training yourself — upload the dataset and watch the model learn! |
+| **⚙️ Watch Training Live** | [Training Space](https://huggingface.co/spaces/YashashMathur/aegis_training) | Real-time metrics — rewards climbing, decisions improving, the model actually learning! |
+| **🎮 Try the Demo** | [Demo Space](https://huggingface.co/spaces/YashashMathur/aegis-demo) | Play with policy violation detection — pick scenarios or enter your own! |
+
+
+**💡 Pro Tip:** Start with the Demo to see what AEGIS does, then dive into Colab to train your own model!
+
+---
+
+## 📁 What's in This Repo
 
-| Resource | Link |
-|----------|------|
-| **Live Environment** | https://huggingface.co/spaces/YashashMathur/aegis_training |
-| **Training Dashboard** | https://yashashmathur-aegis-training.hf.space |
-| **Full Blog Post** | See BLOG.md in this repository |
+| File | What It Does |
+|------|--------------|
+| `train.py` | The main GRPO training script — loads Qwen2.5-7B, runs SFT + GRPO |
+| `openenv.yaml` | OpenEnv manifest — defines observations, actions, rewards |
+| `aegis_training_data_500.json` | **Your dataset** — 500 scenarios with roles, CoT, outputs, labels |
+| `colab_training.ipynb` | Google Colab version — upload this file to Colab to train! |
+| `world_model.py` | WorldModelSimulator — intercepts SQL/shell/code, returns suspicion scores |
+| `memory.py` | MemoryLedger — tracks past violations for compound attack detection |
+| `app.py` | Gradio demo — the fun interactive version for everyone |
+| `BLOG.md` | The full story — why we built this, how it works, what we learned |
 
 ---
 
-## Files
-
-| File | Description |
-|------|-------------|
-| `train.py` | GRPO training script |
-| `openenv.yaml` | OpenEnv framework manifest |
-| `aegis_training_data_500.json` | 500 scenario dataset |
-| `world_model.py` | WorldModelSimulator |
-| `memory.py` | MemoryLedger for compound attack detection |
-| `demo_dashboard.py` | Gradio demo for presentations |
-| `BLOG.md` | Full technical write-up |
+## 🔧 Training on Google Colab
+
+Want to train on Colab? Here's how:
+
+1. **Open the notebook:** https://colab.research.google.com/drive/1hwg_GexKSdTiBY3njPciMUhcvTDvC85w?usp=sharing
+
+2. **Upload `aegis_training_data_500.json`** (from this repo's root folder) when prompted
+
+3. **Run cells** — the notebook handles dependency installation automatically
+
+4. **Watch training** — Colab provides GPU time, you provide the data!
 
 ---