Local Deep Research

Community-submitted benchmark results for Local Deep Research (LDR) on SimpleQA, BrowseComp, and xbench-DeepSearch.

This repository is the source of truth for raw YAML submissions, PR review, CI validation, and leaderboard regeneration. Aggregated leaderboard CSVs are synced automatically to a companion Hugging Face dataset for browsing and discovery:

👉 huggingface.co/datasets/local-deep-research/ldr-benchmarks

Quick contribution path

1. Run your benchmark from the LDR web UI at /benchmark (LDR v0.6.0+)
2. On the Benchmark Results page, click the YAML button next to your completed run — this downloads a pre-filled file
3. Review the file:
   • Fill in any blank hardware / version fields you know
   • For BrowseComp or xbench runs, strip any examples: block (required — these benchmarks prohibit plain-text question/answer sharing; see Benchmark sharing policy below)
4. Open a Pull Request adding the file at:

   results/{dataset}/{strategy}/{search_engine}/{model}_{YYYY-MM-DD}.yaml
   

   • {dataset} — simpleqa, browsecomp, xbench-deepsearch
   • {strategy} — e.g. focused-iteration, source-based, langgraph-agent, or any other LDR strategy (hyphenated form)
   • {search_engine} — e.g. searxng, serper, tavily
5. CI runs automatically on your PR, validating the schema and flagging any issues. A maintainer merges once checks pass.

After merge, GitHub Actions regenerates leaderboards/*.csv and pushes them to the Hugging Face dataset.

Repository layout

results/                        # raw YAML submissions (source of truth)
  simpleqa/
    {strategy}/
      {search_engine}/
        {model}_{date}.yaml
  browsecomp/
    ...
  xbench-deepsearch/
    ...

leaderboards/                   # auto-generated, do not edit by hand
  all.csv
  simpleqa.csv
  browsecomp.csv
  xbench-deepsearch.csv

scripts/
  validate_yamls.py             # schema + policy validator (CI)
  build_leaderboards.py         # YAML → CSV aggregator (CI)
  sync_to_hf.py                 # pushes CSVs + HF README to HF (CI)

.github/workflows/
  validate.yml                  # runs on every PR
  publish.yml                   # runs on merge to main

hf_README.md                    # README shown on the Hugging Face dataset
README.md                       # this file

Result file schema

The schema follows LDR's official benchmark_template.yaml. The easiest way to produce a well-formed file is to use the built-in exporter in the LDR web UI, which generates this format automatically.

# Model Information
model: qwen3.5:9b
model_provider: OLLAMA
quantization:                    # optional, e.g. Q4_K_M

# Search Engine
search_engine: serper
search_provider_version:
average_results_per_query:

# Hardware
hardware:
  gpu:
  ram:
  cpu:

# Benchmark Results
results:
  dataset: SimpleQA              # simpleqa | browsecomp | xbench_deepsearch
  total_questions: 200

  source_based:                  # any LDR strategy key; include as many
    accuracy: "91.2% (182/200)"  # as were actually run
    iterations: 10
    questions_per_iteration: 1
    avg_time_per_question: "1m 18s"
    total_tokens_used:

# Configuration
configuration:
  context_window: 36352
  temperature: 0.7
  max_tokens: 30000
  local_provider_context_window_size: 36352

# Evaluator / Grader
evaluator:
  model: anthropic/claude-3.7-sonnet
  provider: openai_endpoint
  endpoint_url: https://openrouter.ai/api/v1
  temperature: 0

# Versions
versions:
  ldr_version: 1.5.6
  ollama_version:
  searxng_version:

# Test Details
test_details:
  date_tested: 2026-04-06
  rate_limiting_issues:
  search_failures:

# Notes
notes: |
  Any observations about this run.

# Individual examples (optional, SimpleQA only)
# examples:
#   - question: "..."
#     correct_answer: "..."
#     model_answer: "..."
#     result: correct            # correct | incorrect
#     dataset: simpleqa
#     processing_time_seconds: 45.3

Strategy keys under results: are taken from LDR's internal strategy names (e.g. source_based, focused_iteration, langgraph_agent, evidence_based, etc.). Use whichever strategies you actually ran. The path uses the hyphenated form (source-based/, focused-iteration/).

Supported benchmarks & sharing policy

Benchmark	Canonical ID	Path slug	Examples allowed?
SimpleQA	simpleqa	simpleqa	✅ Yes
BrowseComp	browsecomp	browsecomp	❌ No (encrypted, canary-protected)
xbench-DeepSearch	xbench_deepsearch	xbench-deepsearch	❌ No (encrypted)

• SimpleQA is MIT-licensed and openly distributed. Full per-question examples are welcome in submissions.
• BrowseComp is distributed encrypted with a canary string. OpenAI explicitly states that benchmark data should never appear as plain text online. CI will reject any BrowseComp submission containing an examples: block.
• xbench-DeepSearch is also distributed encrypted. The xbench team asks that plain-text questions/answers not be uploaded online. CI will reject any xbench submission containing an examples: block.

⚠️ If you are not sure whether it is safe to share the examples — do not upload them. Summary-only submissions (no examples: block) are the default and preferred format for every benchmark. You lose nothing in the leaderboards by omitting examples.

Adding a new benchmark

When LDR adds support for a new benchmark dataset, update the BENCHMARKS whitelist in both:

• scripts/validate_yamls.py
• scripts/build_leaderboards.py

Keep the canonical_id in sync with LDR's src/local_deep_research/benchmarks/datasets/__init__.py registry.

Considerations for using the data

This is a community-submitted leaderboard so the numbers here are estimates and not ground truth. This section explains how much to trust a result and when a difference between two runs is meaningful.

Self-reported results. CI validates schema and path conventions but cannot verify that a run happened exactly as described. Treat every submission as coming from a good-faith contributor, but apply the guardrails below before drawing conclusions.

Uncertainty scales with sample size. A reported accuracy is a point estimate with a CI that depends on how many questions were tested. The Wilson score interval gives the correct range at 95% confidence: Use at least 100 examples before drawing any conclusions, and 200+ before comparing two configurations.

Small observed differences are likely noise. To reliably detect a real accuracy difference between two configurations (80% statistical power, α = 0.05), each configuration needs roughly:

Difference to detect	Examples needed per config
5 pp (e.g., 85% vs 90%)	~680
10 pp (e.g., 80% vs 90%)	~200
15 pp (e.g., 75% vs 90%)	~90

If the observed gap between two runs is smaller than the margin of error for either run, treat the results as a tie.

Evaluator noise sets a practical floor. The grader LLM mis-grades approximately 1% of responses. This means a 1–2 percentage-point difference is indistinguishable from grader error alone, regardless of sample size. Differences smaller than ~2–3 pp are not actionable.

Cross-run comparison requires identical conditions. Large sample sizes cannot save a comparison where any of the following differ between runs: LDR version, strategy, grader model, random seed, or search engine. Each combination of these factors is a distinct experimental condition. See METHODOLOGY.md for detail and a pre-flight checklist.

Timing is environment-dependent. Compare avg_time_per_question with caution across different hardware and network setups.

Contamination risk. SimpleQA is publicly distributed and model providers may have trained on it. BrowseComp and xbench-DeepSearch mitigate this with encryption and per-question canary strings.

Contributor attribution

Each CSV row includes a contributor column and a contributor_source column indicating where the name came from:

• yaml — the submitter set an explicit contributor: field in the YAML. This wins over everything else.
• git — no explicit field; the aggregator ran git log --follow --diff-filter=A on the result file and used the author of the commit that first added it. Later edits to the file do not overwrite the original contributor.
• "" (empty) — neither a YAML field nor a git history was available (e.g. running outside a git checkout).

If you want credit under a handle other than your git commit name, add contributor: your-handle to the YAML.

Contributors

Thanks to everyone who has contributed benchmark runs:

• LearningCircuit — 9 submissions
• Daniel Petti — 1 submission
• kwhyte7 — 1 submission

See CONTRIBUTORS.md for the full list (auto-generated).

Local development

# Install dependencies (pyyaml is the only hard requirement)
pip install pyyaml

# Run unit tests (uses stdlib unittest, no pytest required)
python -m unittest tests.test_scripts -v

# Validate all submissions
python scripts/validate_yamls.py

# Validate a specific file
python scripts/validate_yamls.py results/simpleqa/source-based/serper/qwen3.5-9b_2026-04-06.yaml

# Regenerate leaderboards locally
python scripts/build_leaderboards.py

# Dry-run the HF sync
python scripts/sync_to_hf.py --dry-run

Setting up the HF sync (maintainers)

1. Create a write token on Hugging Face with access to local-deep-research/ldr-benchmarks
2. Add it as a repo secret on GitHub named HF_TOKEN
3. The publish.yml workflow picks it up automatically on merge to main

Attribution

• SimpleQA © OpenAI, MIT License
• BrowseComp © OpenAI
• xbench © xbench team
• Local Deep Research © the LDR community

Citation

If you use this dataset in published work, please cite both the individual benchmarks (see the Attribution section) and this collection:

@misc{ldr_community_benchmarks,
  title        = {LDR Community Benchmarks},
  author       = {The Local Deep Research community},
  year         = {2026},
  publisher    = {GitHub / Hugging Face},
  howpublished = {\url{https://github.com/LearningCircuit/ldr-benchmarks}}
}

License

TBD. License not yet chosen — likely CC BY 4.0 for the whole repo, or MIT, pending discussion. Individual benchmark datasets (SimpleQA, BrowseComp, xbench) retain their own upstream licenses regardless — see the Attribution section.