olmOCR is a toolkit for converting PDFs and other image-based document formats (such as PNG and JPEG) into clean, readable plain text exported as Markdown. It supports equations, tables, handwriting, and complex formatting, and it can remove headers and footers while producing a natural reading order. It also provides an online demo and a benchmark suite (olmOCR-Bench) for evaluating OCR performance.
Project status
- Actively maintained: The repository shows multiple recent tagged updates (0.4.14 through 0.4.27) and the upstream activity is recent (last push on 2026-03-25), indicating ongoing development rather than maintenance-only.
- Update cadence: There were several updates clustered in mid-to-late January 2026 (0.4.14 to 0.4.21), followed by additional updates in late January (0.4.24, 0.4.25) and a later update in March (0.4.27).
AI summary generated
Recent updates
v0.4.27
Release v0.4.27 contains mostly documentation and maintenance changes, with a few internal refactors. The code diff shows a notable internal reorganization around front matter parsing, moving logic and changing import locations.
v0.4.25
Release v0.4.25 mainly updates the version metadata and makes a small change to the rich PII tagging pipeline. The release notes only mention that the pipeline is faster, with no additional behavioral details.
v0.4.24
Release v0.4.24 bumps the patch version and includes changes described only at a high level as "PII Pipeline fixes" and "Larger timeouts". The code diff shows multiple concrete behavior changes in the PII tagging pipelines, plus changes to Beaker job configuration and multimodal model limits.
v0.4.21
Release v0.4.21 primarily updates the Beaker job submission code paths and the PII tagging pipelines. The release notes do not describe any user-visible schema or workflow behavior changes, but the code diff shows multiple significant modifications to Beaker secret handling and the tagging output schema.
Breakingv0.4.20
Release v0.4.20 appears to be a metadata-only update. The code diff shows only a version increment and corresponding changelog entry, with no functional logic changes.
v0.4.19
v0.4.19 mainly bumps the patch version and changes the behavior of the page backoff retry logic. When maximum backoff attempts are reached, the pipeline now terminates the running process instead of returning a null result.
Breakingv0.4.18
The v0.4.18 release notes do not describe any functional changes, only a version bump commit and a generic “formatting fixes” commit. However, the code diff shows meaningful logic changes in the rendering cache schema and pipeline retry behavior that are not mentioned in the release notes.
Breakingv0.4.16
v0.4.16 primarily introduces changes inside the main document processing pipeline, including a retry/refactor likely aimed at improving extraction robustness. The release notes for this version do not describe any of these behavioral changes.
Featuresv0.4.15
v0.4.15 focuses on operational robustness and viewer/data handling. It adds filesystem operation timeouts in the local work queue backend, refactors stats printing logic, and extends the Dolma viewer to read PDFs from tar.gz archives using a :: path syntax.
BreakingFeaturesv0.4.14
v0.4.14 release notes are essentially empty, with only a version bump and a note that lints were fixed. The actual diff shows meaningful behavioral changes in the PDF processing pipeline, plus updates to benchmark invocation defaults and some README requirements and examples.