Update readme.md

readme.md

@@ -1,32 +1,66 @@
-# HeuristixAI Self-Reflect Qwen 0.5B (v1)
+# HeuristixAI Self-Reflect Qwen 0.5B
 
-This repository contains LoRA adapters trained to induce
-self-reflective reasoning behavior in a compact language model.
+This repository contains LoRA adapters trained to induce **self-reflective reasoning** in a compact language model.
+
+The model learns to follow a structured pattern:
+
+Prompt → Initial Answer → Self-Critique → Revised Answer
+
+---
 
 ## Base Model
+
 Qwen/Qwen2.5-0.5B-Instruct
 
+---
+
 ## Method
-Parameter-efficient fine-tuning (LoRA) on reflection-formatted data:
-Prompt → Initial Answer → Self-Critique → Revised Answer
+
+Parameter-efficient fine-tuning (LoRA) on reflection-formatted data.
+
+Each training example contains:
+
+- Prompt
+- Initial Answer (intentionally imperfect)
+- Self-Critique
+- Revised Answer
+
+This explicitly teaches the model to evaluate and correct its own outputs.
+
+---
+
+## Training Setup
+
+- LoRA rank: 8  
+- Alpha: 16  
+- Dropout: 0.05  
+- Epochs: 3  
+- Learning rate: 2e-4  
+- Sequence length: 512  
+- Quantization: 4-bit NF4  
+- Dataset size: 120 reflection examples  
+- Peak VRAM: ~2.8 GB  
+- Training time: ~20 minutes (GTX 1650)
+
+Only LoRA adapters are provided.
+
+---
 
 ## Capabilities
+
 - Structured reasoning
-- Self-critique behavior
+- Explicit self-critique
 - Reduced hallucination
 - Improved logical consistency
+- Better explanation quality
 
-## Training Setup
-- LoRA r=8, alpha=16, dropout=0.05
-- 4-bit NF4 quantization
-- Dataset size: 120 reflection examples
-- Peak VRAM: ~2.8 GB
-- Training time: ~20 minutes (GTX 1650)
+---
 
 ## Usage
-See reflection_lora_v1_demo.py for example inference.
 
-## License
-Adapters released for research use.
----
-Developed by HeuristixAI.
+See `reflection_lora_v1_demo.py` for inference example.
+
+Install dependencies:
+
+```bash
+pip install -r requirements.txt