Update README.md

README.md

@@ -48,19 +48,19 @@ model-index:
 
 ## ✨ Key Features
 
-- 🎯 **Domain-specific:** Focused exclusively on actuarial and insurance Q&A
-- 📚 **Educational:** Makes complex actuarial terminology accessible for all users
-- 🚀 **Efficient:** Fine-tuned with Unsloth for rapid, scalable training
-- ✨ **Open Source:** Apache 2.0 License; easy to reuse, adapt, remix
-- 🌐 **Widget & Demo:** Integrated as a live demo on [ActuaryEnough](https://actuaryenough.vercel.app)
+- 🎯 **Domain-specific:** Focused exclusively on actuarial and insurance Q&A.
+- 📚 **Educational:** Makes complex actuarial terminology accessible for all users.
+- 🚀 **Efficient:** Fine-tuned with Unsloth for rapid, scalable training.
+- 🔓 **Open Source:** Apache 2.0 License; easy to reuse, adapt, remix.
+- 🌐 **Widget & Demo:** Integrated as a live demo on [ActuaryEnough](https://actuaryenough.vercel.app).
 
 ---
 
 ## 💡 Intended Use Cases
 
-- **Education**: For students and actuaries in training, or for professionals retraining in actuarial language
-- **Translation**: Make practical insurance questions understandable at professional actuarial level
-- **Research**: Support for actuarial research, Q&A, and domain adaptation
+- **Education:** For students and actuaries in training, or for professionals retraining in actuarial language.
+- **Translation:** Make practical insurance questions understandable at professional actuarial level.
+- **Research:** Support for actuarial research, Q&A, and domain adaptation.
 
 ### Examples
 
@@ -74,59 +74,23 @@ Output: "This relates to premium calculation considering risk factors such as ex
 
 ## 📂 Training Data
 
-- **Dataset**: Over 11,000 manually curated actuarial question–answer pairs
-- **Topics**: Life and non-life insurance, risk, regulation, reserves, actuarial mathematics
-- **Language**: English
+- **Dataset:** Over 11,000 manually curated actuarial question–answer pairs.
+- **Topics:** Life and non-life insurance, risk, regulation, reserves, actuarial mathematics.
+- **Language:** English.
 
 ---
 
-## 🔬 Training Procedure & Metrics
-
-- **Base Model**: unsloth/gemma-3-270m-it
-- **Epochs**: ~51 epochs (visible from screenshots, final point in `train/epoch`)
-- **Steps**: over 68,000 global steps (`train/global_step` chart)
-- **Training Loss**:  
-  - Starts around **2.2**  
-  - Smoothly drops, converges to **~1.4** at the final epoch ([see "train/loss" graph])
-- **Learning Rate**:  
-  - Decays linearly from **8e-7** down to near zero ([see "train/learning_rate" graph])
-- **Gradient Norm**:  
-  - Usually oscillates between **5** and **15** ([see "train/grad_norm" graph])
-- **Hardware**:  
-  - NVIDIA GeForce RTX 3090 (24GB VRAM)  
-  - 16 physical/32 logical core CPU  
-  - 94GB RAM  
-  - CUDA 12.8, Linux 6.10
+## 📊 Training Statistics
 
----
-
-## ![Train Loss Curve](attached_image:1)
-
-The curve shows rapid loss reduction in the first epochs, then stable convergence, confirming healthy optimization.
-
----
-
-## ![Learning Rate Schedule](attached_image:2)
-
-Steady linear learning rate decay visible throughout the training cycles.
-
----
-
-## ![Gradient Norms](attached_image:3)
-
-Gradient norms remain well controlled, with only rare spikes, indicating stable training.
-
----
-
-## ![Training Epoch Progress](attached_image:4)
-
-The model was trained for over 50 epochs (as shown on the epoch chart).
-
----
-
-## ![Global Steps](attached_image:5)
-
-A steady climb to over 68,000 update steps during training.
+| Metric             | Value / Range          | Notes                                      |
+|--------------------|-----------------------|--------------------------------------------|
+| Epochs             | ~51                   | Reached at end of training                 |
+| Global Steps       | >68,000               |                                            |
+| Initial Train Loss | ~2.2                  | At start                                   |
+| Final Train Loss   | ~1.4                  | At end                                     |
+| Learning Rate      | 8e-7 → ≈0             | Linear decay throughout training           |
+| Gradient Norm      | 5 – 15                | Generally stable with rare spikes          |
+| Hardware           | RTX 3090, 16-core CPU | 24GB VRAM, 94GB RAM, CUDA 12.8, Linux 6.1  |
 
 ---
 
@@ -146,10 +110,10 @@ pydantic==2.11.7
 
 ## ⚠️ Limitations & Ethics
 
-- **No pricing or decision support:** For education and inspiration only, not for real insurance contracts
-- **Not a substitute for an actuary:** Always consult professionals for real-world decisions
-- **Coverage:** Designed and tested specifically for the insurance/actuarial domain
-- **Training data bias:** Outputs may reflect source content
+- **No pricing or decision support:** For education and inspiration only, not for real insurance contracts.
+- **Not a substitute for an actuary:** Always consult professionals for real-world decisions.
+- **Coverage:** Designed and tested specifically for the insurance/actuarial domain.
+- **Training data bias:** Outputs may reflect source content.
 
 ---