Update README.md

README.md

@@ -28,6 +28,12 @@ This is a **free preview** of the full **CYB003 — Synthetic Malware Behaviour
 full dataset** at identical schema, family/tier distribution, and statistical
 fingerprint, so you can evaluate fit before licensing the full product.
 
+> 🤖 **Trained baseline available:**
+> [**xpertsystems/cyb003-baseline-classifier**](https://huggingface.co/xpertsystems/cyb003-baseline-classifier)
+> — XGBoost + PyTorch MLP for 10-class malware execution-phase prediction,
+> group-aware split by sample, multi-seed evaluation (accuracy 0.905 ± 0.010),
+> honest disclosure of which tasks need the full dataset.
+
 | File                       | Rows (sample) | Rows (full)   | Description                                    |
 |----------------------------|---------------|---------------|------------------------------------------------|
 | `environment_profiles.csv` | ~100            | ~3,200        | Endpoint environment configurations            |
@@ -53,6 +59,30 @@ stacks with calibrated detection/evasion outcomes, covering:
 - **Outcome modelling**: AV signature detection, EDR behavioural detection,
   sandbox evasion success, family attribution confidence
 
+## Trained Baseline Available
+
+A working baseline classifier trained on this sample is published at
+**[xpertsystems/cyb003-baseline-classifier](https://huggingface.co/xpertsystems/cyb003-baseline-classifier)**.
+
+| Component | Detail |
+|---|---|
+| Task | 10-class malware execution-phase classification |
+| Models | XGBoost (`model_xgb.json`) + PyTorch MLP (`model_mlp.safetensors`) |
+| Features | 69 (after one-hot encoding); pipeline included as `feature_engineering.py` |
+| Split | **Group-aware by sample_id** — train/val/test samples disjoint |
+| Validation | Single seed + multi-seed aggregate across 10 seeds |
+| Demo | `inference_example.ipynb` — end-to-end copy-paste |
+| Headline metrics | XGBoost: accuracy 0.905 ± 0.010, macro ROC-AUC 0.975 ± 0.002 (multi-seed) |
+
+The model card documents an honest finding worth knowing before licensing:
+**malware-family classification is at majority baseline on the sample's 100
+samples** (a sample-size constraint, not a method failure — the full
+280k-row dataset has ~5,600 samples and supports family classification
+properly). The baseline pivots to **execution-phase prediction**, which is
+strongly learnable on the sample data (91% accuracy, ROC-AUC 0.98, stable
+across 10 seeds) and is itself a real SOC use case for dynamic-analysis
+and EDR phase tagging.
+
 ## Calibrated Benchmark Targets
 
 The full product is calibrated to 12 benchmark metrics drawn from
@@ -126,8 +156,11 @@ event log and endpoint environment schemas respectively.
 
 ## Suggested Use Cases
 
+- Training **malware execution-phase classifiers** —
+  [worked example available](https://huggingface.co/xpertsystems/cyb003-baseline-classifier)
 - Training **malware family classifiers** (9-class with realistic class
-  imbalance and family-specific feature distributions)
+  imbalance and family-specific feature distributions — full dataset
+  recommended for adequate per-class sample size)
 - **Threat actor attribution** modelling (4-tier classification)
 - **EDR detection benchmarking** — packed vs unpacked, signature vs
   behavioural, fileless vs binary
@@ -163,6 +196,10 @@ y_tier = summaries["threat_actor_tier"]
 y_detected = (samples["av_detected"] | samples["edr_detected"]).astype(int)
 ```
 
+For a worked end-to-end example with execution-phase classification,
+group-aware splitting, and feature engineering, see the inference notebook
+in the [baseline classifier repo](https://huggingface.co/xpertsystems/cyb003-baseline-classifier/blob/main/inference_example.ipynb).
+
 ## License
 
 This **sample** is released under **CC-BY-NC-4.0** (free for non-commercial