Update README.md

README.md

@@ -1,36 +1,34 @@
-nnul/sqlchat: A Conversational AI for SQL Generation
+# `nnul/sqlchat`: A Conversational AI for SQL Generation
 
-This repository contains sqlchat, a powerful and efficient language model designed specifically for Text-to-SQL tasks. It can understand natural language questions and database schemas to generate accurate SQL queries, including complex statements for creating and managing tables (Data Definition Language).
+This repository contains `sqlchat`, a powerful and efficient language model designed specifically for **Text-to-SQL** tasks. It can understand natural language questions and database schemas to generate accurate SQL queries, including complex statements for creating and managing tables (Data Definition Language).
 
-This model is provided as a standalone 4-bit quantized model, optimized for easy deployment and high-performance, low-resource inference. It was built using the Unsloth library to ensure maximum speed and memory efficiency.
+This model is provided as a standalone 4-bit quantized model, optimized for easy deployment and high-performance, low-resource inference. It was built using the [Unsloth](https://github.com/unslothai/unsloth) library to ensure maximum speed and memory efficiency.
 
-Model Capabilities
+## Model Capabilities
 
-Natural Language to SQL: Translates complex English questions into executable SQL queries.
+*   **Natural Language to SQL:** Translates complex English questions into executable SQL queries.
+*   **Schema-Aware:** Understands `CREATE TABLE` contexts provided in the prompt to generate relevant queries.
+*   **DDL Generation:** Capable of generating `CREATE TABLE` statements, including constraints like `PRIMARY KEY` and `FOREIGN KEY` relationships.
+*   **Complex Query Logic:** Successfully handles `JOIN`s, aggregations (`COUNT`, `MAX`), and sorting (`ORDER BY ... LIMIT`).
 
-Schema-Aware: Understands CREATE TABLE contexts provided in the prompt to generate relevant queries.
+## How to Use
 
-DDL Generation: Capable of generating CREATE TABLE statements, including constraints like PRIMARY KEY and FOREIGN KEY relationships.
+The easiest way to use `sqlchat` is with the Unsloth library, which will ensure you get the best performance.
 
-Complex Query Logic: Successfully handles JOINs, aggregations (COUNT, MAX), and sorting (ORDER BY ... LIMIT).
-
-How to Use
-
-The easiest way to use sqlchat is with the Unsloth library, which will ensure you get the best performance.
-
-Prerequisites
+### Prerequisites
 
 First, install the necessary libraries.
 
-Generated bash
-pip install unsloth
+```bash
+pip install "unsloth[conda]"
 pip install "torch>=2.3.1"
+```
 
-Running Inference
+### Running Inference
 
 Here is a simple, reusable Python script to run inference with the model.
 
-Generated python
+```python
 import torch
 from unsloth import FastLanguageModel
 from transformers import TextStreamer
@@ -92,14 +90,11 @@ CREATE TABLE students (student_id INTEGER PRIMARY KEY, student_name VARCHAR(255)
 CREATE TABLE courses (course_id INTEGER PRIMARY KEY, course_title VARCHAR(255));
 """
 )
-IGNORE_WHEN_COPYING_START
-content_copy
-download
-Use code with caution.
-Python
-IGNORE_WHEN_COPYING_END
-Expected Output
-Generated code
+```
+
+### Expected Output
+
+```
 User Instruction: Which department has the most number of employees?
 
 Model Output:
@@ -113,22 +108,19 @@ Model Output:
 ---------------------------------
 CREATE TABLE student_enrollment (student_id INTEGER, course_id INTEGER, PRIMARY KEY (student_id, course_id), FOREIGN KEY (student_id) REFERENCES students(student_id), FOREIGN KEY (course_id) REFERENCES courses(course_id));
 ---------------------------------
-IGNORE_WHEN_COPYING_START
-content_copy
-download
-Use code with caution.
-IGNORE_WHEN_COPYING_END
-Performance
+```
+
+## Performance
 
-The model was benchmarked on an NVIDIA A40 GPU. In a batch-processing scenario, it achieves a throughput of ~55-70 tokens/second. Single-prompt latency is well within real-time requirements for interactive applications.
+The model was benchmarked on an NVIDIA A40 GPU. In a batch-processing scenario, it achieves a throughput of **~55-70 tokens/second**. Single-prompt latency is well within real-time requirements for interactive applications.
 
-Peak VRAM Usage (Inference): ~6.1 GB
+*   **Peak VRAM Usage (Inference):** ~6.1 GB
 
-Prompt Template
+## Prompt Template
 
 To get the best results, your prompts should follow this structure:
 
-Generated code
+```
 <|im_start|>system
 You are a helpful assistant that generates SQL queries based on natural language questions and database schemas.<|im_end|>
 <|im_start|>user
@@ -138,8 +130,4 @@ You are a helpful assistant that generates SQL queries based on natural language
 ### Context:
 {The CREATE TABLE statements for the relevant tables}<|im_end|>
 <|im_start|>assistant
-IGNORE_WHEN_COPYING_START
-content_copy
-download
-Use code with caution.
-IGNORE_WHEN_COPYING_END
+```