vania-janet/multiturn-rag-retrieval-data

MT-RAG Benchmark - Retrieval Results

This dataset contains experimental results from the Multi-Turn RAG (MT-RAG) benchmark focusing on retrieval tasks across multiple domains.

Dataset Description

Competition: MT-RAG Benchmark - Task A (Retrieval)
Date: January 2026
Domains: CLAPNQ, CLOUD, FIQA, GOVT

Contents

1. Baseline Results with Ground Truth Rewrites

Directory: submissions/baselines_rewrite/

Results for 5 retrieval models using query rewrites provided in the original dataset:

• BM25: Traditional sparse retrieval
• SPLADE: Learned sparse retrieval (best sparse model)
• BGE-1.5: Dense retrieval (BAAI/bge-large-en-v1.5)
• BGE-M3: Multi-modal dense retrieval
• Voyage-3: Commercial dense retrieval API

20 experiments total (5 models × 4 domains)

2. Hybrid Retrieval Results

Directory: submissions/hybrid/

Results combining sparse and dense methods using Reciprocal Rank Fusion (RRF):

• SPLADE + Voyage-3: For CLAPNQ and GOVT (strongest domains)
• SPLADE + BGE-1.5: For CLOUD and FIQA (cost-effective)

Both with and without query rewrites.

8 experiments total (4 configurations × 2 rewrite variants)

File Structure

submissions/
├── baselines_rewrite/
│   ├── A2_baseline_bm25_rewrite/
│   │   ├── clapnq/
│   │   │   ├── metrics.json
│   │   │   └── retrieval_results.jsonl
│   │   ├── cloud/
│   │   ├── fiqa/
│   │   └── govt/
│   ├── A2_baseline_splade_rewrite/
│   ├── A2_baseline_bge15_rewrite/
│   ├── A2_baseline_bgem3_rewrite/
│   └── A2_baseline_voyage_rewrite/
│
├── hybrid/
│   ├── hybrid_splade_voyage_norewrite/
│   ├── hybrid_splade_voyage_rewrite/
│   ├── hybrid_splade_bge15_norewrite/
│   └── hybrid_splade_bge15_rewrite/
│
└── RESULTS_SUMMARY.md

Metrics

Each experiment includes metrics.json with:

• nDCG @ 5, 10, 20, 100
• Recall @ 5, 10, 20, 100
• MAP @ 5, 10, 20, 100
• Precision @ 5, 10, 20, 100

Best Results (nDCG@10)

Domain	Configuration	Score
CLAPNQ	SPLADE + Voyage-3 (Rewrite)	0.56266
GOVT	SPLADE + Voyage-3 (Rewrite)	0.53445
CLOUD	SPLADE + BGE-1.5 (Rewrite)	0.44028
FIQA	SPLADE + BGE-1.5 (Rewrite)	0.40589

Average: 0.48582

Key Findings

1. Ground truth rewrites are effective: +9% to +26% improvement over last-turn queries
2. SPLADE is the best sparse retriever: Consistent performance across all domains (avg 0.457)
3. Hybrid methods outperform individual retrievers: +3% to +10% improvement
4. Domain-specific optimization matters: Voyage-3 for strong domains, BGE-1.5 for weaker/cost-sensitive
5. BGE-M3 underperforms with rewrites: Should be avoided in rewrite scenarios

Retrieval Results Format

Each retrieval_results.jsonl contains one JSON object per query:

{
  "task_id": "clapnq_123",
  "question": "How do I configure SSL?",
  "contexts": [
    {
      "document_id": "doc123_chunk5",
      "score": 0.85,
      "text": "To configure SSL..."
    },
    ...
  ],
  "Collection": "clapnq",
  "turn_id": 3
}

Citation

If you use this dataset, please cite:

@dataset{mt_rag_retrieval_results_2026,
  title={MT-RAG Benchmark Retrieval Results},
  author={Your Name},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/vania-janet/multiturn-rag-retrieval-data}
}

License

Apache 2.0

Additional Information

For methodology details, see RESULTS_SUMMARY.md in the dataset.