The numbers are public, and so is the corpus that produced them.
Datalab scores 94.1 on OmniParseBench, our open benchmark of 16,288 tests. Your team can rerun every test instead of taking our word for it.
USE CASE · MODEL TRAINING DATA
Clean training data, at the throughput your pipeline needs.
When you build a training corpus, the parse is where signal gets lost — tables flatten, equations break, scripts mangle. Datalab turns crawls, archives, and papers into clean, layout-preserving markdown at scale.
TRUSTED BY
Five million pages of scientific literature, parsed in days.
CASE · 04
Radical AI
Read study →"We were quite blown away by the quality of extraction, especially on the type of complex academic papers that we were passing in."
- 5M+ PAGES INDEXED, IN DAYS
- 2–3/s PAGES PER SECOND, SUSTAINED
- 16 wks TO MATERIALS PERFORMANCE GOALS
From crawl to corpus, on a pinned pipeline.
from datalab_sdk import DatalabClient
client = DatalabClient()
PIPELINE_ID = "p_training_corpus_v3" # published, immutable version
for doc in crawl("s3://crawls/2026-q1/"):
execution = client.run_pipeline(PIPELINE_ID, file_path=doc)
execution = client.get_pipeline_execution(execution.execution_id, max_polls=300)
result = client.get_step_result(execution.execution_id, step_index=0)
if result.parse_quality_score >= 3.0:
write("s3://training/2026-q1/", result.markdown)
# > capacity for 1B+ pages in a day
# > Parse Quality Score filter applied (>= 3.0)
# > pipeline version pinned - same configuration across re-runsWhy teams prefer Datalab for model training.
When a document type breaks the parse, the next release fixes it.
If documents in your corpus parse poorly, our research team takes them back to Chandra and ships a fix in the next release.
We tune to the documents you actually care about.
Scientific notation, multilingual scripts, microfilm scans, equation-heavy PDFs — the parts where parsing usually drops are what we improve.
Run it wherever your training data has to live.
- Managed cloud
Start with an API key — nothing to host or operate.
Sign up → - EU data residency
Run in-region, with no egress to US infrastructure.
Sign up → - Your VPC
Runs inside your own AWS, GCP, or Azure account.
Talk to sales → - On-prem & air-gapped
Fully offline, the same model weights, dedicated support.
Talk to sales →
SOC 2 Type II · BAA available View our trust center →
SCALE FOR TRAINING
- 1B+pages in a day
- 94.1OmniParseBench · Datalab accurate
- 76k+open-source stars · Marker, Surya, Chandra, and Indra
START
Stop training on dirty markdown.
Free tier with up to $20 in credits per month — no card required.