Datasets:

sightandcite
/

SpectRAG-eval-results

license: mit
language:
  - en
pretty_name: SpectRAG Evaluation Results
size_categories:
  - n<1K
task_categories:
  - question-answering
tags:
  - rag
  - retrieval-augmented-generation
  - multi-hop-qa
  - graph-rag
  - evaluation
  - benchmark-results
configs:
  - config_name: retrieval_ranking
    data_files: data/retrieval_ranking.csv
    default: true
  - config_name: answer_accuracy
    data_files: data/answer_accuracy.csv
  - config_name: accuracy_by_question_type
    data_files: data/accuracy_by_question_type.csv
  - config_name: index_cost
    data_files: data/index_cost.csv
  - config_name: pairwise_win_rates
    data_files: data/pairwise_win_rates.csv

SpectRAG Evaluation Results

Aggregate evaluation results from the paper SpectRAG: Weighted Chunk Graphs for Token-Budgeted Context Selection in Multi-Hop Retrieval-Augmented Generation.

SpectRAG is compared with NaiveRAG, Hybrid (BM25 + dense, RRF), HyDE and LightRAG on MultiHopRAG and four UltraDomain corpora (Agriculture, CS, Legal, Mix). Every system uses the same generator (gpt-4o-mini, temperature 0), so differences come from retrieval.

📄 Paper: arXiv:2610.36991 · 🤗 Paper page
💻 Code: github.com/SightAndCite/SpectRAG
🎬 Visual walkthrough: sightandcite/SpectRAG Space

Subsets

Config	Source	Rows	What it contains
`retrieval_ranking`	Table 6	5	MultiHopRAG gold-evidence recall (overall, @2, @5, @10) and MRR per system
`answer_accuracy`	Table 7	5	Exact match, token F1 and answer containment on 2,556 MultiHopRAG questions
`accuracy_by_question_type`	Table 8	20	Exact match per question type (comparison, inference, temporal, null) and system
`index_cost`	Table 4	10	Graph nodes, edges, edges/node and graph file size per corpus, SpectRAG vs LightRAG
`pairwise_win_rates`	Tables 9 & 10	160	gpt-4o pairwise judge win rates, per edge-weight config, baseline, domain and dimension

All values are percentages except nodes, edges, edges_per_node, graph_file_mb and node_reduction_vs_lightrag.

Usage

from datasets import load_dataset

ranking = load_dataset("sightandcite/SpectRAG-eval-results", "retrieval_ranking", split="train")
wins = load_dataset("sightandcite/SpectRAG-eval-results", "pairwise_win_rates", split="train").to_pandas()

# SpectRAG's average diversity win rate against each baseline
print(wins.query("domain == 'Average' and dimension == 'diversity'")
          [["signal_config", "baseline", "spectrag_win_rate"]])

Column notes

pairwise_win_rates: in each row baseline_win_rate + spectrag_win_rate = 100. A spectrag_win_rate above 50 means the judge preferred SpectRAG. Answer order was swapped to control for position bias. domain = "Average" rows are the averages reported in the paper. edge_weights lists the six edge-signal weights used for that configuration.
accuracy_by_question_type: null questions have no answerable gold evidence, so a higher score means correct abstention.
retrieval_ranking: the paper notes that NaiveRAG's chunk-level retrieval does not map cleanly onto the graph-node evidence-attribution scheme, so treat its retrieval figures with care. Its generation-level results in answer_accuracy are directly comparable.

Limitations

These are aggregate results transcribed from the paper's tables (Tables 9 and 10 were parsed programmatically from the PDF and every row was checked to sum to 100). They don't include per-question predictions or judge rationales. Pairwise judgments come from an LLM judge (gpt-4o) on a sample of questions per UltraDomain domain, so read small differences with caution.

Citation

@misc{sahinoglu2026spectragweightedchunkgraphs,
      title={SpectRAG: Weighted Chunk Graphs for Token-Budgeted Context Selection in Multi-Hop Retrieval-Augmented Generation},
      author={Zafer Sahinoglu and Abdulazeez Mohammed and Buse Yaren Kazangirler and Cahit Berkay Kazangirler},
      year={2026},
      eprint={2610.36991},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.36991},
}