Upload README.md with huggingface_hub

README.md

@@ -29,13 +29,18 @@ Unlike any traditional astrological software, this model was trained on 100,000
 3. It cross-references the 12 outputs against a massive database of 100+ encoded professions.
 4. It predicts the *best profession* and **deduces the hidden astrological birth hour** simultaneously.
 
-## Model Details
+## Technical Deep-Dive: Saju-Attention Mechanism
 
-- **Architecture:** Saju-Attention (Custom PyTorch Multihead Attention + SentenceTransformer Semantics)
-- **Base LLM Embedding:** `sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2`
-- **Training Data:** 100,000 real-world biographical profiles extracted from Wikipedia.
-- **Certified Precision:** 88.32% Cosine Similarity on unseen validation data.
-- **Language:** English semantics (handles basic Multilingual inference via the base MiniLM).
+Traditional Saju assumes a static relationship between the Four Pillars. **Saju 4.0** treats these pillars as tokens in a semantic sequence:
+1. **Embedding Layer:** Year, Month, Day, and Time pillars are converted into 384-dimensional dense vectors using a pre-trained `paraphrase-multilingual-MiniLM-L12-v2` transformer.
+2. **Multi-Head Attention:** A custom PyTorch layer allows the "Tiger" of the Year to pay "attention" to the "Rabbit" of the Day, discovering hidden synergistic correlations that simple linear models miss.
+3. **Semantic Anchoring:** Instead of predicting a fixed category, the model projects the Saju-Context into the *same vector space* as professions. This allows for **Open-Vocabulary Inference**—you can ask the model about "Quantum Physicist" or "Content Creator" even if those weren't explicitly used during training.
+
+## The MIL Moonshot: Reverse Rectification
+The most innovative feature is how we handled the lack of birth times in our 100k training records:
+- **Challenge:** Wikipedia bio data almost never includes birth hours.
+- **Solution:** We implemented **Multiple Instance Learning (MIL)**. During every training step, the model evaluates all 12 potential time pillars for an individual. It selects the one that yields the highest similarity to the known profession (the "witness") and updates its attention weights based on that optimized temporal path. 
+- **Result:** The model learned not just what jobs people have, but *which astrological hours* are most probable for high-achievers in specific fields.
 
 ## How to play with it locally
 
@@ -43,19 +48,22 @@ You can use the model immediately using the standard PyTorch loading mechanism a
 
 ```bash
 # Clone the repository
-git clone https://huggingface.co/[YOUR-USERNAME]/saju-v4-universal
-cd saju-v4-universal
+git clone https://huggingface.co/srSergio/saju
+cd saju
 
-# Run the inference engine for a specific date (Format: YYYY-MM-DD)
-python saju_inference.py 1993-02-15
+# Run the inference engine (Format: YYYY-MM-DD)
+python saju_inference.py 1998-01-07
 ```
 
-If you know your exact hour of birth, you can pass it to bypass the MIL multiversal rectification and get your deterministic Top 10 career paths:
+If you know your exact hour of birth (0-24), you can pass it to get a deterministic Top 10 career ranking:
 ```bash
-# For a person born at 14:00 (2 PM)
-python saju_inference.py 1993-02-15 14
+python saju_inference.py 1998-01-07 22
 ```
 
+## Hardware & Ethics
+- **Inference:** Lightweight, runs on any modern CPU in < 1 second.
+- **Ethics:** This is a probabilistic statistical model based on historical biographical data. It should be used for vocational inspiration and entertainment, not for critical life-path decisions.
+
 ## How to use the Python Class directly
 
 ```python