Feat: Upload fine-tuned Requin-Agent security model with calibrated temperature

README.md

@@ -14,13 +14,10 @@ language:
 pipeline_tag: text-classification
 ---
 
-# Requin-Agent Security Decision Engine
-
-System 1 calibrated decision model for zero-latency threat triage:
-- **OWASP LLM01**: Prompt Injection & Jailbreak detection
-- **Web Exploits**: SQL Injection, XSS, Command Injection, SSRF
-- **MITRE ATT&CK**: Privilege Escalation, Credential Access, Data Exfiltration
-- **Latency**: ~35ms per request
-
-## Usage
+# Requin-Agent Security Decision Engine (Fine-tuned Laya)
 
+Fast, calibrated System 1 Risk Investigator for zero-latency threat triage:
+- **Trained on 88,000+ security cases** (OWASP LLM01, Web Exploits, MITRE ATT&CK).
+- **Latency**: ~32ms per request on Tesla T4.
+- **Calibrated ECE**: < 0.08 with L-BFGS Temperature Scaling.
+- **Routing Actions**: `allow` (safe), `invoke_llm` (medium-risk investigation), `escalate_human` (critical/unseen incident), `block` (high-threat exploit).

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bd770a73af7b9c6c847c3978d3b8c5d81611bc565d813f3d12b8e302f51ec268
+oid sha256:2c4d7f9443e02bb05044b1fe78f620aa19b14a0b0984b4b860f2c95140ce317c
 size 643835524

rl_agent_config.json

@@ -11,9 +11,9 @@
   "amp_dtype": "bf16",
   "model_name": "requin-security-decision-engine",
   "temperature": [
-    0.8024038076400757,
-    0.6024259924888611,
-    0.6070500612258911
+    0.8575455546379089,
+    0.5979790091514587,
+    0.5893478393554688
   ],
   "temperature_by_options": {},
   "training": {