Update README.md

README.md

@@ -1,26 +1,62 @@
+---
+language:
+- en
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-classification
+tags:
+- cybersecurity
+- phishing-detection
+- email-security
+- text-classification
+- onnx
+- int8
+- modernbert
+base_model: answerdotai/ModernBERT-base
+base_model_relation: finetune
+widget:
+- text: "Subject: Security Alert\n\nBody:\nYour account has been locked. Please reply with your password immediately to restore access."
+  example_title: "Phishing-like email"
+- text: "Subject: Team Lunch Reminder\n\nBody:\nReminder that the team lunch is tomorrow at 12:30 PM in the office kitchen."
+  example_title: "Benign email"
+---
+
 # ModernBERT Phishing Detector
 
-This project fine-tunes ModernBERT-base for phishing email detection.
-
-Artifacts:
-- `models/modernbert_phish/` -> PyTorch fine-tuned model
-- `onnx/modernbert_phish/model.onnx` -> ONNX export
-- `onnx/modernbert_phish/model.int8.onnx` -> quantized ONNX export
-- `models/modernbert_phish/calibration.json` -> score calibration values
-
-Scoring:
-- margin = phish_logit - safe_logit
-- probability = sigmoid(coef * margin + intercept)
-- score_0_10 = round(10 * probability)
-- score_1_10 = max(1, round(10 * probability))
-
-Suggested UI colors:
-- 0-2 = green
-- 3-5 = yellow
-- 6-7 = orange
-- 8-10 = red
-
-Evidence extraction:
-- split the email into sentences
-- score each sentence independently
-- return the highest-risk sentence as the explanation text
+## Model description
+
+This model is a fine-tuned **ModernBERT-base** binary sequence classifier for **phishing email detection**. It takes a full email as input text and predicts whether the email is **safe** or **phishing**.
+
+The training backbone is **`answerdotai/ModernBERT-base`**, and the final release includes:
+- a fine-tuned PyTorch checkpoint
+- an ONNX export
+- a quantized INT8 ONNX export
+- a calibration file for mapping logits to a user-facing phishing score
+
+## Intended use
+
+This model is intended for:
+- phishing detection in email text
+- mobile or backend inference through ONNX Runtime
+- UI risk scoring, such as a **0–10** or **1–10** phishing scale
+- evidence extraction via sentence-level rescoring
+
+This model is **not** intended for:
+- malware analysis
+- attachment sandboxing
+- URL detonation
+- image/PDF threat inspection
+- general prompt-injection detection
+- fully explainable token-level rationale extraction
+
+## Inputs
+
+The model expects a single text string representing the email content.
+
+Example format:
+
+```text
+Subject: Urgent Account Notice
+
+Body:
+Your account has been locked. Please reply with your password immediately to restore access.