Update README.md

README.md

@@ -28,6 +28,13 @@ Dataset** product. It contains roughly **~10% of the full dataset** at
 identical schema, actor-tier distribution, and statistical fingerprint, so
 you can evaluate fit before licensing the full product.
 
+> 🤖 **Trained baseline available:**
+> [**xpertsystems/cyb007-baseline-classifier**](https://huggingface.co/xpertsystems/cyb007-baseline-classifier)
+> — XGBoost + PyTorch MLP for **3-tier insider threat type classification**
+> (the README's stated headline use case), group-aware split by incident,
+> multi-seed evaluation (ROC-AUC 0.961 ± 0.007), honest leakage audit of
+> tier-correlated volume features.
+
 | File                          | Rows (sample) | Rows (full)   | Description                                  |
 |-------------------------------|---------------|---------------|----------------------------------------------|
 | `org_topology.csv`            | ~240           | ~2,400        | Department / org structure registry          |
@@ -59,6 +66,32 @@ defender modeling, covering:
 - **Sabotage outcomes** — destructive insider actions distinct from
   exfiltration
 
+## Trained Baseline Available
+
+A working baseline classifier trained on this sample is published at
+**[xpertsystems/cyb007-baseline-classifier](https://huggingface.co/xpertsystems/cyb007-baseline-classifier)**.
+
+| Component | Detail |
+|---|---|
+| Task | **3-class insider threat type classification** (the README's headline use case) |
+| Models | XGBoost (`model_xgb.json`) + PyTorch MLP (`model_mlp.safetensors`) |
+| Features | 28 (after one-hot encoding); pipeline included as `feature_engineering.py` |
+| Split | **Group-aware by incident_id** — train/val/test incidents disjoint |
+| Validation | Single seed + multi-seed aggregate across 10 seeds |
+| Demo | `inference_example.ipynb` — end-to-end copy-paste |
+| Headline metrics | XGBoost: accuracy 0.855 ± 0.012, macro ROC-AUC 0.961 ± 0.007 (multi-seed); MLP slightly outperforms (acc 0.869, AUC 0.966 at seed 42) |
+
+This is the **second XpertSystems baseline to ship the dataset's stated
+headline use case** (after CYB005). CYB007's 500-incident sample is large
+enough that tier attribution learns honestly under group-aware splitting,
+with no oracle features and very tight multi-seed std.
+
+**Important schema note for buyers:** the dataset README documents a
+4-tier scheme including `compromised_account`, but the sample contains
+only 3 of those 4 tiers. The baseline trains on what exists in the
+sample. See the baseline model card for the full list of schema
+discrepancies between README and data.
+
 ## Calibrated Benchmark Targets
 
 The full product is calibrated to 12 benchmark validation tests drawn from
@@ -134,7 +167,9 @@ and department registry schemas respectively.
 
 ## Suggested Use Cases
 
-- Training **insider threat classifier** models (4-tier actor attribution)
+- Training **insider threat classifier** models (3-tier actor attribution
+  on the sample, 4-tier on the full product) —
+  [worked example available](https://huggingface.co/xpertsystems/cyb007-baseline-classifier)
 - **Data exfiltration detection** modelling — DLP signal calibration
 - **UEBA effectiveness benchmarking** — graduated 8-tier defender maturity
 - **HR-signal correlation** — disgruntlement, resignation, performance
@@ -160,16 +195,20 @@ topology     = pd.read_csv("org_topology.csv")
 enriched = trajectories.merge(incidents, on=["incident_id", "actor_id"],
                               how="left", suffixes=("", "_summary"))
 
-# 4-class threat-type classification target
-y_tier = incidents["threat_type_tier"]
+# 3-class threat-type classification target (sample contains 3 of 4 README tiers)
+y_tier = trajectories["actor_threat_type"]
 
 # Binary exfiltration-success target
-y_exfil = incidents["exfiltration_success_flag"]
+y_exfil = incidents["exfiltration_successes"] > 0
 
-# Binary collusion target
-y_collusion = incidents["collusion_flag"]
+# Binary coordinated-incident target
+y_coord = incidents["coordinated_incident_flag"]
 ```
 
+For a worked end-to-end example with insider-threat tier classification,
+group-aware splitting, and feature engineering, see the inference notebook
+in the [baseline classifier repo](https://huggingface.co/xpertsystems/cyb007-baseline-classifier/blob/main/inference_example.ipynb).
+
 ## License
 
 This **sample** is released under **CC-BY-NC-4.0** (free for non-commercial