Merge pull request #27 from MIT-Emerging-Talent/qwen

Qwen2.5 Pipeline

Author: AseelOmer

.gitignore

@@ -0,0 +1,57 @@
+<!-- markdownlint-disable MD013 -->
+# QwenModel: Qwen 2.5 + RAG + Recursive Editing
+
+This project experiments with the open-sourced model **Qwen 2.5** (2.5B parameters) combined with **Retrieval-Augmented Generation (RAG)** and **recursive editing** to test its performance across different prompt categories from the Apollo 11 dataset.
+
+## Project Structure
+
+* **main.ipynb** – The core notebook where Qwen 2.5 is set up, RAG is integrated, and recursive editing is implemented.
+* **output.md** – Contains the final prompts and responses generated through recursive editing, along with tracked energy and CO₂ emission metrics.
+
+## Workflow Overview
+
+### 1. Prompt Setup
+
+The notebook defines three structured prompts for iterative refinement:
+
+* **Draft Prompt** – Produces the initial answer.
+* **Critic Prompt** – Analyzes weaknesses in the draft.
+* **Refiner Prompt** – Improves the answer based on the critique.
+
+This setup enables controlled recursive editing, especially useful for summarization, creative writing, and paraphrasing.
+
+### 2. RAG Pipeline
+
+* Connected to the Hugging Face API to generate **embeddings**.
+* Built a **vector index** from the Apollo 11 dataset.
+* Prepared a **retriever** to supply relevant context for Qwen.
+
+### 3. Qwen Integration
+
+A custom function handles:
+
+* Retrieval of context
+* Construction of Qwen API calls
+* Application of recursive editing logic (3 iterations)
+
+### 4. Energy Tracking
+
+Each query is wrapped with **CodeCarbon** to record:
+
+* CPU usage
+* Energy consumption
+* CO₂ emissions
+
+All tracked metrics appear in `output.md`.
+
+## Output
+
+The final markdown file contains:
+
+* The prompts used
+* Recursive-editing improved final responses
+* Per-question energy and emission statistics
+
+## Purpose
+
+This setup provides a compact testing environment to evaluate Qwen 2.5’s performance with RAG and recursive editing, enabling comparison with larger or alternative models.