Datasets:
Languages:
English
Size:
10K<n<100K
Tags:
retrieval-augmented-generation
rag
causal-reasoning
multi-hop-qa
semantic-drift
context-window-poisoning
License:
Add README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
task_categories:
|
| 4 |
+
- question-answering
|
| 5 |
+
- text-retrieval
|
| 6 |
+
language:
|
| 7 |
+
- en
|
| 8 |
+
tags:
|
| 9 |
+
- retrieval-augmented-generation
|
| 10 |
+
- rag
|
| 11 |
+
- causal-reasoning
|
| 12 |
+
- multi-hop-qa
|
| 13 |
+
- semantic-drift
|
| 14 |
+
- context-window-poisoning
|
| 15 |
+
- benchmark
|
| 16 |
+
- evaluation
|
| 17 |
+
pretty_name: VORTEXRAG Benchmark Evaluation Data
|
| 18 |
+
size_categories:
|
| 19 |
+
- 10K<n<100K
|
| 20 |
+
---
|
| 21 |
+
|
| 22 |
+
# VORTEXRAG Benchmark Evaluation Dataset
|
| 23 |
+
|
| 24 |
+
Benchmark evaluation data, results, and per-layer ablation scores for the VORTEXRAG 7-layer causal RAG framework.
|
| 25 |
+
|
| 26 |
+
## Description
|
| 27 |
+
|
| 28 |
+
This dataset contains:
|
| 29 |
+
1. **Main benchmark results** comparing VORTEXRAG against 7 baseline systems across NQ, HotpotQA, MuSiQue, and 2WikiMultiHopQA
|
| 30 |
+
2. **Full ablation study** showing per-layer contribution (configs A through H)
|
| 31 |
+
3. **Per-dataset breakdown** showing performance on 6 evaluation datasets
|
| 32 |
+
4. **Domain preset evaluation** for all 11 domain configurations
|
| 33 |
+
5. **Hyperparameter sensitivity** sweep data for τ and θ_CPG
|
| 34 |
+
6. **Latency breakdown** per layer on A100 GPU
|
| 35 |
+
|
| 36 |
+
## Datasets Used for Evaluation
|
| 37 |
+
|
| 38 |
+
| Dataset | Type | Questions | Hops | Source |
|
| 39 |
+
|---------|------|-----------|------|--------|
|
| 40 |
+
| NaturalQuestions | Open-domain QA | 7,842 | 1–2 | Wikipedia Dec 2018 |
|
| 41 |
+
| HotpotQA | Multi-hop QA | 7,405 | 2 | Wikipedia 10K docs |
|
| 42 |
+
| MuSiQue | Multi-hop QA | 2,417 | 2–4 | Wikipedia filtered |
|
| 43 |
+
| 2WikiMultiHopQA | Multi-hop QA | 12,576 | 2 | Wikipedia + Wikidata |
|
| 44 |
+
| LegalBench | Legal QA | 1,200 | 1–3 | US federal case law |
|
| 45 |
+
| MedQA | Medical QA | 1,273 | 1–2 | PubMed abstracts |
|
| 46 |
+
|
| 47 |
+
## Main Results
|
| 48 |
+
|
| 49 |
+
| System | EM | F1 | Faithfulness | SDR | CPR | Latency |
|
| 50 |
+
|--------|----|----|--------------|-----|-----|---------|
|
| 51 |
+
| Naive RAG | 61.2 | 68.4 | 0.71 | — | — | 120ms |
|
| 52 |
+
| BM25+Rerank | 59.8 | 66.1 | 0.69 | — | — | 95ms |
|
| 53 |
+
| HyDE | 64.1 | 71.8 | 0.74 | 12% | 8% | 340ms |
|
| 54 |
+
| CRAG | 66.9 | 74.3 | 0.78 | 31% | 19% | 290ms |
|
| 55 |
+
| Self-RAG | 68.4 | 75.9 | 0.81 | 35% | 24% | 410ms |
|
| 56 |
+
| FiD | 63.5 | 70.2 | 0.73 | 8% | 5% | 280ms |
|
| 57 |
+
| FLARE | 65.7 | 72.9 | 0.75 | 14% | 10% | 320ms |
|
| 58 |
+
| **VORTEXRAG** | **74.8** | **82.6** | **0.94** | **61%** | **71%** | **185ms** |
|
| 59 |
+
|
| 60 |
+
## Links
|
| 61 |
+
- 📄 Paper: https://doi.org/10.5281/zenodo.20285144
|
| 62 |
+
- 💻 GitHub: https://github.com/vignesh2027/VORTEXRAG
|
| 63 |
+
- 🤗 Space: https://huggingface.co/spaces/vigneshwar234/VORTEXRAG
|
| 64 |
+
|
| 65 |
+
**Author:** Vignesh L | **License:** MIT | **Version:** v2.0
|