You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
ControlMind is a MinerU-based scientific document intelligence project for the 2026 MinerU Data Intelligence Challenge. It connects three independent tracks into one auditable chain: scientific corpus construction, autonomous data processing, and source-grounded RAG.
3
+
**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.**
| Track 2: Data Agent | 14-intent Agent protocol with scheduling, logs, fallback, replay, and validation |[track2_agent_report.md](docs/submissions/track2_agent_report.md)|
34
-
| Track 3: Medical RAG | Source-grounded medical literature RAG with Chinese Ask, claim support, refusal boundary, and local deployment |[track3_medical_rag_report.md](docs/submissions/track3_medical_rag_report.md)|
*ControlMind system architecture: a single RTX 5090 runs the full pipeline — MinerU parsing, tri-engine inference (API / Ollama / vLLM), and multi-index RAG.*
| Track 3 Medical RAG | 97 parsed PMC papers, 3,348 medical chunks, FAISS/BM25/vision indexes |
51
-
| Chinese Ask | BGE-M3 fixed trace with full claim support and citation coverage in saved evaluation |
52
-
| Local-first boundary | Medical chunks, indexes, QLoRA data, and RAG context stay local by default |
23
+
| Track | What You Get | One Command |
24
+
|:---|:---|:---|
25
+
|**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")`|
26
+
|**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`|
27
+
|**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`|
53
28
54
-
All quantitative claims in the submission reports point back to files, commands, or hashes in [DATA-TRACE.md](docs/submissions/shared/DATA-TRACE.md).
29
+
---
55
30
56
31
## Quick Start
57
32
58
-
Use an existing `myenv` environment when available.
59
-
60
-
```powershell
61
-
conda run -n myenv python demo/cli/controlscidemo all --quick
62
-
```
63
-
64
-
Run per-track quick views:
65
-
66
-
```powershell
67
-
conda run -n myenv python demo/cli/controlscidemo track1 --quick
68
-
conda run -n myenv python demo/cli/controlscidemo track2 --quick
69
-
conda run -n myenv python demo/cli/controlscidemo track3 --quick
33
+
```bash
34
+
pip install -r requirements.txt
35
+
pip install -e .
36
+
controlmind doctor
70
37
```
71
38
72
-
Run reviewer-oriented minimal verification:
39
+
Load the Sci-Align benchmark dataset:
73
40
74
-
```powershell
75
-
.\run_reviewer_demo.ps1 -Track All -SkipApiHealth
76
-
```
77
-
78
-
Run JSON-oriented checks. Use `--output` when saving files on Windows; it lets the Python CLI write UTF-8 JSON directly and avoids shell redirection encoding issues.
41
+
```python
42
+
from datasets import load_dataset
79
43
80
-
```powershell
81
-
conda run --no-capture-output -n myenv python -m controlsci.cli doctor --output _scratch/doctor.json
82
-
conda run --no-capture-output -n myenv python -m controlsci.cli track2 validate --artifact all --output _scratch/track2_validate.json
> **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.
96
58
97
-
```powershell
98
-
npm install -g ./npm/controlmind
99
-
controlmind wrapper-doctor
100
-
controlmind track2 validate --artifact all
101
-
```
102
-
103
-
The npm package is only a thin launcher. It locates the repository, selects `CONTROLMIND_PYTHON`, `conda run -n myenv python`, or system Python, and forwards commands to `python -m controlsci.cli`.
Every quantitative claim in the reports points back to source files, commands, or hashes in [`DATA-TRACE.md`](docs/submissions/shared/DATA-TRACE.md).
149
107
150
-
The public cloud demo is for public or sanitized examples. Private documents, medical chunks, indexes, model adapters, and RAG contexts are handled through local/private paths.
108
+
---
151
109
152
-
## Reproducibility Boundary
110
+
## Public Entry Points
153
111
154
-
The recommended verification path uses packaged public artifacts and local indexes. arXiv/PMC online downloading remains available as a corpus expansion capability, but it is intentionally not a prerequisite for minimal validation because public services may apply rate limits, browser checks, or short-lived cookies. Local run outputs should go under `_scratch/`; that directory is ignored and can be deleted or regenerated without affecting the audited source data.
For the full layered policy, see [REPRODUCIBILITY.md](REPRODUCIBILITY.md). In short: smoke checks and minimal real-chain validation are designed for a public checkout; report-level conclusions are auditable through `docs/submissions/data_trace_bundle/`; full-scale rebuilds may require GPU, network access, and external model/API credentials.
119
+
---
157
120
158
121
## Repository Map
159
122
160
123
```text
161
-
benchmark/ Track 1 benchmark and Track 2 Agent artifacts
0 commit comments