Published results
Explore scientific PDF-to-JSON extraction across four public arXiv papers, with standalone facts, equations, visual findings, and named entities. The corpus spans model architecture, mathematical finance, astrophysics, and language-model experiments: 74 pages in total. Both archived runs used DeepSeek V4.1 Flash through OrcaRouter and completed every phase.
These outputs let you inspect the scientific records available before building a retrieval index or knowledge graph. The use cases explain how to use them in RAG, GraphRAG, and semantic search applications.
Several small and mid-sized open-weight language models were evaluated on this workflow. The aim was to process papers at scale while keeping cost per page low; larger models may improve extraction quality. Only the strongest model on this corpus, DeepSeek V4.1 Flash, is retained in the public showcase.
Results by paper
Model cells contain facts / equations / visuals and link to the full document JSON.
| Paper | Pages | DeepSeek T=0 | DeepSeek T=1 |
|---|---|---|---|
| GQA | 7 | 72 / 4 / 7 | 75 / 1 / 7 |
| Dynamic AMM fees | 23 | 177 / 47 / 23 | 180 / 35 / 24 |
| Stellar XUV flux | 23 | 326 / 24 / 18 | 240 / 23 / 22 |
| Algospeak | 21 | 168 / 13 / 21 | 169 / 14 / 19 |
| Run | Facts | Equations | Visuals | Candidates | Main | Secondary | Completed documents |
|---|---|---|---|---|---|---|---|
| DeepSeek T=0 | 743 | 88 | 69 | 872 | 311 | 360 | 4 |
| DeepSeek T=1 | 664 | 73 | 72 | 400 | 183 | 152 | 4 |
Entities are named scientific methods, systems, or concepts. Candidate counts show what the page pass found; main and secondary counts show what document selection retained. Both runs completed all required phases. Generated content is uncorrected; counts measure output volume, not accuracy or recall. Parser warnings remain visible in the raw request artifacts.
DeepSeek V4.1 Flash at temperature 0 preserved more equations and produced fewer document-dependent statements on this corpus. This observation is specific to the measured model and corpus.
The outputs still require entity normalization, cross-page deduplication, and citation-graph construction before large-scale indexing.
Raw outputs
Each run archive contains its documents, prompts, complete provider responses, checkpoints, GROBID TEI, configuration, and request statistics.
See cost and timing, reproduction instructions, and source attribution.