ControlMind

A MinerU-powered scientific document intelligence system: 500-question cross-modal benchmark, 14-intent data agent, and local-first medical RAG — all from raw PDFs.

Chinese version | CC-BY-4.0

Track 1  Sci-Align  —  4-dimension control-science evaluation benchmark
Track 2  Data Agent —  14-intent autonomous corpus agent with 4-path scheduling
Track 3  Medical RAG —  local-first evidence-grounded clinical literature Q&A

---

ControlMind system architecture: a single RTX 5090 runs the full pipeline — MinerU parsing, tri-engine inference (API / Ollama / vLLM), and multi-index RAG.

---

What This Project Does

Track	What You Get	One Command
Sci-Align	A 500-question 4-dimension benchmark (A: Concept Recall, B: Multi-step Reasoning, C: Condition Sensitivity, D: Open Design) with 9-model leaderboard and full source traceability. Loadable via load_dataset().	load_dataset("MorningStar0709/control-sci-corpus")
Data Agent	A 14-intent autonomous agent that searches arXiv, parses PDFs with MinerU, audits cross-modal alignment, builds benchmarks, evaluates models, and self-corrects on failure — with unified logging and checkpoint recovery.	controlmind track2 validate --artifact all
Medical RAG	A local-first evidence Q&A system over 97 parsed PMC papers, with IMRAD-aware chunking, hybrid FAISS+BM25 retrieval, Chinese-to-English query bridging, visual injection, and safety-refusal boundaries.	controlmind track3 eval --case-set zh_ask

---

Quick Start

pip install -r requirements.txt
pip install -e .
controlmind doctor

Load the Sci-Align benchmark dataset:

from datasets import load_dataset

core = load_dataset("MorningStar0709/control-sci-corpus", "core", split="train")
print(len(core))       # 500
print(core[0]["question"])

Run per-track quick checks:

controlmind track1 validate --sample 4
controlmind track2 validate --artifact all
controlmind track3 eval --case-set zh_ask

Windows users: prepend conda run -n myenv python -m controlsci.cli if not using pip install -e .. PowerShell scripts (run_reviewer_demo.ps1, run_frontend.ps1) are also provided.

---

Local Demo

.\run_frontend.ps1 -StartBackend

Starts the backend API at http://127.0.0.1:17001 and the frontend at http://127.0.0.1:3000. Requires Python (FastAPI) and Node.js.

---

Key Results

Metric	Value
Documents parsed	362 (23 textbooks + 339 arXiv papers)
Structured chunks	28,514
LaTeX formulas extracted	253,012
Image-formula co-occurrence pairs	4,996 (9,207 audit judgments)
Benchmark questions	500 (A/B/C/D = 125 each, 14 sub-domains)
Models evaluated	9
PMC medical papers	97 parsed, 3,348 medical chunks
QLoRA fine-tuning	4B/9B variants, perplexity-probed

---

Leaderboard — ControlSci Sci-Align Benchmark

Rank	Model	Overall	A: Concept	B: Reasoning	C: Sensitivity	D: Design
1	MiMo-v2-flash	0.647	0.610	0.606	0.636	0.736
2	DeepSeek-v4-flash	0.632	0.634	0.631	0.714	0.550
3	Qwen3.5-9B	0.625	0.569	0.610	0.662	0.659
4	DeepSeek-v4-pro	0.619	0.627	0.590	0.742	0.514
5	MiniMax-M2.5-highspeed	0.602	0.638	0.519	0.624	0.626
6	MiniMax-M2.7-highspeed	0.574	0.605	0.485	0.612	0.593
7	MiMo-v2.5-pro	0.539	0.595	0.523	0.602	0.436
8	MiMo-v2-pro	0.514	0.638	0.490	0.560	0.369
9	MiMo-v2.5	0.440	0.608	0.466	0.528	0.156

All scores verified by LLM-as-Judge with cross-validation. Full results and analysis in benchmark/eval/results/.

---

Reports

Each track has a companion technical report with detailed methodology, experiments, and traceable evidence:

Track	Report	Key Evidence
Track 1 Sci-Align	track1_sci_align_report.md	500-question schema, 9-model leaderboard, QLoRA results
Track 2 Data Agent	track2_agent_report.md	14-intent protocol, dry-run logs, failure recovery cases
Track 3 Medical RAG	track3_medical_rag_report.md	97 PMC papers, Chinese Ask traces, MedBench comparison

Every quantitative claim in the reports points back to source files, commands, or hashes in DATA-TRACE.md.

---

Public Entry Points

Item	Link
Cloud Demo	demo.askiler.com (code: ControlMind@2026)
HuggingFace Dataset	MorningStar0709/control-sci-corpus
Reproducibility Guide	REPRODUCIBILITY.md
Evidence Bundle	docs/submissions/data_trace_bundle/

---

Repository Map

Source & CLI
  controlsci/                  Python package (controlmind CLI)
  benchmark/                   Benchmark engine, Agent orchestration, datasets
  benchmark/dataset/           Core/full/schema JSON, multimodal index
  benchmark/eval/              Evaluation, leaderboard & medical RAG scripts
  tools/                       MinerU utilities & analysis scripts
  npm/controlmind/             Optional Node.js CLI launcher

Data & Corpus
  corpus/chunks/               28,514 structured chunks from 362 documents
  data/sources_medical/        Medical corpus, chunks, FAISS/BM25 indexes
  data/sources_flywheel/       Agent flywheel papers & parse results

Reports & Pipeline
  docs/submissions/            Technical reports, DATA-TRACE, evidence bundle
  _final_submission_by_track/  Per-track final submission packages
  pipeline/                    Corpus construction pipeline scripts
  notebooks/                   Colab demo & E3b training screenshot
  starboard/                   Local & cloud demo frontend (Next.js)

---

License & Data Boundary

• This project is released under CC-BY-4.0. See LICENSE.
• Public PMC/arXiv source documents retain their original licenses and attribution.
• Patient-level private data is not included.
• Cloud demo inputs are limited to public or sanitized materials; medical chunks, indexes, and RAG contexts are local-first by default.

---

Citation

@misc{controlmind2026,
  title        = {ControlMind: MinerU-based Scientific Document Intelligence for Sci-Align, Data Agent, and Medical RAG},
  author       = {MorningStar},
  year         = {2026},
  howpublished = {\url{https://github.com/MorningStar0709/control-sci}},
  note         = {ControlSci benchmark released under CC-BY-4.0}
}