Update README.md

README.md

@@ -1,64 +1,29 @@
----
-license: mit
----
-
-# SQL Socratic Models
-
-This repository contains fine-tuned large language models for **Socratic SQL instruction** in higher education, focusing on guiding learners through SQL concepts using structured reasoning rather than providing direct solutions.
-
-## Models
-- phi3_rq4
-- qwen25
-- gemma2
-
 ## Method
 
-Our approach is designed to support **conceptual learning in STEM education** through Socratic interaction:
-
-- **Phase 1 (Data Construction):**  
-  SQL instruction data is augmented with Socratic prompts emphasizing:
-  - Question decomposition  
-  - Conceptual hints  
-  - Guided reasoning steps  
-
-- **Phase 2 (Fine-Tuning):**  
-  We apply full fine-tuning (FFT) on small, open-source LLMs with **pedagogical constraints** that explicitly discourage direct answer generation and instead promote:
-  - Conceptual scaffolding  
-  - Incremental reasoning  
-  - Learner-centered guidance  
+Our approach is structured in three phases to support Socratic SQL instruction for higher education.
 
-- **Phase 3 (Evaluation):**  
-  Models are evaluated using:
-  - **BERTScore** for semantic alignment with expected reasoning  
-  - **ROUGE-L** to measure and control **answer leakage** (i.e., avoidance of direct SQL solutions)
+### Phase 1: SQL-Instruct Corpus Construction
 
-## Contributions
-- Fine-tuning across multiple architectures (Phi-3, Qwen2.5, Gemma2) for **instructional SQL reasoning**
-- Development of **Socratic SQL prompting framework** for higher education contexts
-- Evaluation of models on their ability to generate **guidance without revealing final answers**
-- Ablation study identifying factors that enable LLMs to mimic effective instructors through:
-  - Misconception-aware feedback  
-  - Iterative questioning  
-  - Structured reasoning support  
+We construct **SQL-Instruct**, a domain-specific Socratic instruction corpus, by mining high-quality interactions from Stack Overflow. This platform provides a rich source of real-world misconceptions, debugging challenges, and conceptual difficulties encountered by both students and practitioners, making it well-suited for training models that emphasize understanding over code replication.
 
-## Task
+To ensure data quality, we filter SQL-tagged questions based on community impact. The resulting dataset reflects substantial engagement, with a cumulative reach of approximately **1.27 billion views** and an average of **128,535 views per question**. For each selected instance, we extract:
+- The core problem description  
+- User-provided SQL attempts (when available)  
+- Executable SQL blocks from the accepted solution  
 
-Given a natural language SQL question, the model generates:
+This process yields **9,916 unique questions**, which are then transformed into Socratic instructional data using GPT-4o. We leverage GPT-4o for its strong reasoning capabilities to generate **pedagogical hints and guided reasoning steps**, ensuring that the dataset emphasizes conceptual scaffolding rather than direct answers.
 
-1. Socratic reasoning steps  
-2. Conceptual hints and guiding questions  
-3. Intermediate decomposition of the problem  
+The dataset is intentionally skewed toward higher cognitive complexity, with:
+- **8,604 intermediate-level questions**  
+- **629 advanced-level questions**  
 
-**The model does NOT produce the final SQL query**, ensuring alignment with instructional use in higher education settings.
+Additionally, we identify a subset of **531 debugging tasks**, enabling models to learn how to guide students through error identification and correction in SQL queries.
 
-This design supports:
-- Active learning  
-- Conceptual understanding of SQL  
-- Integration of database concepts into broader STEM curricula  
+The corpus spans a wide range of SQL topics, with particular emphasis on:
+- JOIN operations  
+- Aggregation and grouping  
+- Query optimization and performance  
 
-## Usage
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
+By selecting questions with a **median Stack Overflow score of 27**, we ensure that the underlying solutions—and therefore the derived instructional signals—are technically reliable.
 
-model = AutoModelForCausalLM.from_pretrained("sriram882004/SQL-Socratic-Models/phi3_rq4")
-tokenizer = AutoTokenizer.from_pretrained("sriram882004/SQL-Socratic-Models/phi3_rq4")
+This corpus serves as the foundation for training models that prioritize **Socratic reasoning, misconception-aware feedback, and conceptual understanding** over direct SQL solution generation.