Re-OCR'ing the 1771 Britannica with Baidu Unlimited-OCR-2 at $0.0012/page
vanstriendaniel · x · 2026-10-27
HF staff member davanstrien re-OCR'd the 1771 Encyclopaedia Britannica with Baidu's new Unlimited-OCR-2. Key points: it reads many pages in one request (16-page batches), returning cross-page sentences as continuations; cost was $3.40 for 2,724 pages ($0.0012/page, down from $5 in Feb); requires a ≥48GB GPU since a 16-page request holds 31k tokens of KV cache, with rtx-pro-6000 the fastest per dollar; it misreads the long s (ſ) as 'f', so GLM-OCR still gets more words right on 1771 print. The uv-script is open-sourced in the uv-scripts/ocr repo on HF, taking Buckets/folders of page images and outputting parquet.
More from coding & agent
- Judgment Labs launches publicly: monitoring and eval layer for long-horizon agents — testingcatalog · 2026-10-28
- OSU NLP Group Wins Three Paper Awards Same Day at EMNLP 2026 and NeurIPS Workshop — hhsun1 · 2026-10-28
- PPCC26 demos praised: real orgs like NFL, real code live in VS Code, no fake companies — donasarkar · 2026-10-28
- Five principles for curating LLM agent evals: diverse, realistic, hard, discriminative, high-quality — cwolferesearch · 2026-10-28
- figwig ships formal Kaitai Struct spec, auto-generate readers/writers in any language — jmschreiber91 · 2026-10-28
- Harmonic launches Aristotle, an AI theorem prover that formally verifies C and Verilog code — xennygrimmato_ · 2026-10-28