Paper2Profile

One bucket, two parts:

• papers/ — source PDFs (100) + MinerU-2.5 VLM parses (markdown with HTML table grids, content_list.json with bbox/page_idx, pre-cropped block images) + manifest.jsonl.
• result/ — structured, paper-only, verifiable-by-construction extractions (66 papers): per-paper 3-topic CSVs (components_architecture, accuracy_efficiency[_traced], computational), combined/*_ALL.csv, and case_studies/ (reliability evidence pack).

Pipeline: PDF → MinerU 2.5 VLM → clean table grids → hybrid LLM structure-ID + deterministic grid-read → CSVs. The LLM only identifies table structure; deterministic code reads grid[row][col], so every value traces to an exact cell (with table,row,col coords). Unstated computational fields are not_reported (never fabricated).

Reliability (grid-oracle over accuracy cells): 100% dataset-correct, 0 miss, ~82% full (dataset+model); schema-valid 100%; 0 fabricated numbers. See result/case_studies/README.md.

Reproduce

The pipeline lives in the parse_extract/ package of the code repo. Install + run:

# EXTRACT-only (no GPU) — re-run extraction on already-parsed papers:
pip install openai==2.44.0 pyyaml==6.0.3

# FULL parse+extract (CUDA-13 GPU):
conda env create -f parse_extract/environment.yml
conda activate paper2profile
uv pip install -r parse_extract/requirements.txt --index-strategy unsafe-best-match --torch-backend=cu130

# start the extract LLM server, then run the stage:
GPU=0 bash rightsizing/engine/serve_qwen.sh &                                   # Qwen3-14B on :8001
GPU=1 bash parse_extract/run.sh --config parse_extract/configs/full.yaml \
      --steps parse,extract,verify,assemble                                     # -> data/extracted/<arxiv>/*.csv

Models: opendatalab/MinerU2.5-Pro-2605-1.2B (parse VLM) · Qwen/Qwen3-14B (extract structure-ID). Verified installable + runnable from a clean environment.