Update README.md

README.md

@@ -1,32 +1,27 @@
---- 
-library_name: transformers
-tags:
-- qwen3
-- guard
-- safety
-- powershell
-license: apache-2.0
-metrics:
-- accuracy
-- f1
-base_model: Qwen/Qwen3Guard-Stream-0.6B
----
-
-# Qwen3Guard PowerShell Production (Checkpoint 2500)
-
-This model is a fine-tuned version of `Qwen3ForGuardModel` designed for safety moderation, specifically tailored for PowerShell content.
+
+# Codeguard-Stream for PowerShell
+
+Codeguard is a model family based on the Qwen3Guard (primarily stream) models.
+This model is the first variant of a series of models and serves as a proof of concept.
+
+Motivation: As the usage of language models for code generation is increasing, and specifically agentic code generation softwarem we need to ensure that the code that is generated is safe to execute.
+In theory, one should never allow a language model to execute code that could in any way be malicious.
+But in practice, people sacrifice their security for the sake of efficiency, exposing themselves to immense risk if a malicious string of code is executed.
+The code guard  models  is an effort to  add another layer of security on top of  code agents  so that if  malicious code  or  vulnerable code  is  detected 
+the generation is immediately interrupted to prevent execution of said code.
+
+We do not recommend to use this model as a replacement of common sense OPSEC.
+But we believe that this is an important step of research to arrive at a technical setup that can maximize efficiency while minimizing the risk of harm.
+
 
 ## Model Details
 
 - **Model Type**: Qwen3 Guard Stream
 - **Architecture**: `Qwen3ForGuardModel`
-- **Language**: Multilingual (119 languages), specialized for PowerShell
 - **License**: Apache 2.0
 
 ## Training Information
 
-The model was fine-tuned with the following parameters:
-
 - **Epochs**: ~1.55
 - **Global Steps**: 2500
 - **Best Loss**: 0.0777 (at step 1800)
@@ -36,9 +31,6 @@ The model was fine-tuned with the following parameters:
     - **F1 Safe**: 98.66%
     - **F1 Unsafe**: 97.91%
 
-## Usage
-
-This model is designed to be used with the `transformers` library for real-time safety moderation.
 
 ### Example Code