CoreLabel Labelix
The workspace every annotator sees, whatever the data type. Text, image, video, audio, DICOM, RLHF, and the relations that hold an agentic trajectory together — one workspace, kept current, not a fork we walked away from.
A forked tool that isn't maintained is a liability.
Vendors that fork an open-source annotation tool and stop merging inherit its bugs without its fixes. Ours is a full source fork of Label Studio, and it's currently at the same release upstream is — kept current through real, ongoing merges, not frozen at the version we started from.
That matters because this is the workspace behind almost everything we deliver that isn't a robotics job: pre-labeled images and video, entity extraction on text, transcribed and diarized audio, medical imaging bound to clinical ontologies, and the RLHF workflows that train a reward model.
A specific, checkable claim
Most forks drift. Ours has been merged forward release over release, with shared history preserved — which is the difference between "we copied this once" and "we keep this current."
Agent trajectories, labeled step by step
Captured from a live project: 100 agent trajectories, each pre-labeled by a model before a human reviews it.
- Every step is a region — a tool call and its result, labeled on its own.
- A fixed error vocabulary — correct, wrong_tool, wrong_params, redundant_retry, hallucinated_result, unsafe_action, needs_review.
- Model first, human second — a model's pre-labels sit beside the annotator's for comparison.
What Runs Through This Workspace
The engine underneath our Core Services — and part of how our agentic pillar labels relations between steps.
Image & video pre-label
SAM2 and Grounding DINO auto-segment, detect, and classify before a human ever opens the task — the starting point behind our Image and Video Labeling services.
Text & NER pre-label
spaCy, BART, and DeBERTa handle entity extraction, relation tagging, and classification ahead of review.
Audio annotation
Transcription, speaker diarization, sentiment tagging, and event detection built on WhisperX.
DICOM medical annotation
Pixel-level segmentation bound to SNOMED, ICD-10, and RADLEX ontology — used under the same governance approach described on our Data Governance page.
RLHF workflows
SFT pairs, red-teaming prompts, reward-model evaluation, and Constitutional AI critique-and-revise loops — the workspace behind our dedicated RLHF Tuning page.
Agentic relations
Multi-agent handoffs and long-horizon memory recall are labeled here using this workspace's own relation tags — the connective tissue for our Agentic AI Annotation pillar.
Everything labeled here lands in Datum.
Whatever an annotator produces in this workspace — a bounding box, a transcribed span, a preference ranking — flows into the same schema, cleaning pipeline, and export engine as work done in Robotix. One data platform underneath both studios, so the format you get back doesn't depend on which tool did the labeling.
Want to see the workspace before you commit a batch?
We'll walk you through it on a call — the same view your annotators use.