qsardor/viorra-admissions-essays

🎓 VIORRA Admissions Essays Dataset (v2.0.0)

<p align="center"> <img src="https://img.shields.io/badge/Dataset%20Version-2.0.0-blueviolet?style=for-the-badge" alt="Dataset Version"> <img src="https://img.shields.io/badge/License-CC%20BY--NC%204.0-red?style=for-the-badge" alt="License"> <img src="https://img.shields.io/badge/Total%20Essays-615-green?style=for-the-badge" alt="Total Essays"> <img src="https://img.shields.io/badge/RAG%20Index-TurboVec-orange?style=for-the-badge" alt="TurboVec Index"> </p>

⚖️ License: CC BY-NC 4.0 — Free for academic, research, and personal use. Attribution required. Commercial distribution or usage is strictly prohibited.

VIORRA Admissions Essays is a highly curated, multi-category dataset of successful personal statements that resulted in admission to elite global universities (Ivy League, Oxbridge, and Top 20 schools). It is explicitly designed to train, evaluate, and ground conversational EdTech AI mentors using advanced Retrieval-Augmented Generation (RAG) and post-training alignment.

🌟 What's New in v2.0.0?

• —🚀 Expanded Corpus: Upgraded from 59 to 615 high-signal documents, covering both paired (essay + official admissions feedback) and unpaired successful drafts.
• —🌍 Global Formats: Added native support for UK/UCAS (academic style), UCLA PIQ (short prompt response), and MBA Insead (professional career statements) alongside the traditional US Common App (narrative style).
• —⚡ TurboVec Integration: Compiled natively as a pre-indexed 4-bit quantized SIMD vector database (viorra_index.tv), dropping all heavy FAISS dependencies for near-instant boot and sub-millisecond search latencies on local CPU/GPU hardware.

📊 Dataset Structure

The dataset contains a diverse mix of application formats to train models to differentiate narrative, academic, and professional writing styles:

JSON Schema (verbatim keys in JSONL files)

{
  "messages": [
    {"role": "system", "content": "You are VIORRA, an elite college admissions mentor..."},
    {"role": "user", "content": "[Student Essay Draft]"},
    {"role": "assistant", "content": "The essay demonstrates strong self-reflection...", "reasoning": "<think>First, analyze the narrative hook. The transition is abrupt...</think>"}
  ]
}

⚡ Performance Benchmarks (Local CPU/GPU)

The dataset's pre-compiled RAG binaries (viorra_index.tv and viorra_corpus.pkl) are optimized for sub-millisecond vector lookups utilizing TurboVec (4-bit quantized SIMD):

• —Index Boot Time: ~115.4 ms (FastEmbed initialization + index mapping)
• —Search Latency: ~0.0400 ms (Averaged over 200 queries, L2 distance mapping)
• —Active RAM Footprint: <1.7 GB (Fully loaded cache ready for real-time inference)

🛠️ Getting Started / Local RAG Setup

To load and query the pre-compiled RAG database natively without FAISS:

import os
import pickle
import numpy as np
import turbovec
from fastembed import TextEmbedding

# 1. Load the pre-compiled corpus and native index
index = turbovec.TurboQuantIndex.load("viorra_index.tv")
index.prepare()  # Warm up SIMD search caches

with open("viorra_corpus.pkl", "rb") as f:
    corpus_texts, corpus_feedback = pickle.load(f)

# 2. Embed user query using MiniLM-L6
embedder = TextEmbedding(model_name="sentence-transformers/all-MiniLM-L6-v2")
query = "A draft about Czech heritage and exploring history under Communism."
query_emb = np.array(list(embedder.embed([query])), dtype=np.float32)

# 3. Fast Vector Query
distances, indices = index.search(query_emb, 2)
for i, idx in enumerate(indices[0]):
    print(f"Match #{i+1}: {corpus_texts[idx][:200]}...")

⚖️ License, Attribution & Ethics

1. 1.Strictly Non-Commercial: This dataset is created for educational research, automated grading validation, and fine-tuning AI agents to critique writing. Any commercial licensing must be approved in writing by the authors.
2. 2.No Plagiarism Support: This dataset should never be used to train models designed to write essays for students. AI-generated admissions essays are easily detected and damage educational integrity.
3. 3.PII Redaction: All sensitive personal identifying information (PII) has been thoroughly sanitized and redacted.

✍️ Citation

If you use this dataset in your academic publications or open-source software, please cite it as follows:

@dataset{viorra_admissions_essays_2026,
  author    = {Sardor},
  title     = {VIORRA Admissions Essays Dataset v2.0.0},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/qsardor/viorra-admissions-essays},
  license   = {CC BY-NC 4.0}
}

Created & Curated by:

• —Sardor & The Antigravity AI Agent Team (Built via human-agent pair programming).