Data Workshop
Turn real behaviour into data
Your production traces are already the best training data you have. Overmind turns them into training and eval data: selected, audited, and redacted, with a profile that configures the training run that follows.
From traces to training-ready
Your agents already generate the best training data you will ever have. The Workshop makes it usable.
Build datasets from real runs
Select production traces and promote them into a dataset. Filter by agent, eval score, or time window so only the runs worth learning from make it through. Every example keeps a link back to its source trace.
Audited by detectors, then by an agent
Deterministic checks catch duplicates, malformed turns, token overflow, and PII in seconds. Then a coding agent reads the corpus for what rules miss: label conflicts, answer leakage, degenerate outputs. Every fix it proposes is verified in a sandbox first, and applying one is a single click that lands as a versioned commit.
The profile configures the training run
The Workshop profiles shape, balance, and token distributions, then turns them into a recommended training configuration: model tier, epochs, learning rate, and adapter settings, with cost and duration estimated before you commit. PII is detected and redacted so what leaves the Workshop is safe to train on.
What you get
Curation, validation, and profiling in one place, wired into evals and training.
Curation
Trace selection
Bulk-select runs filtered by agent, score, and time. Choose the agent surface or raw model call.
Provenance tracking
Every example links back to its source trace, so any training row can be audited to the run that made it.
Dataset history
Every applied fix is a commit that records the score it produced. Restore any point in the history.
Validation
Structure checks
Validates turn order, tool-call shape, declared-tool usage, and token overflow before training can start.
Agent-fit checks
Tool overlap, input-key match, and provenance checks so the dataset matches the agent it will train.
Quality pillars
Weighted scores across integrity, signal, coverage, uniqueness, safety, and fit. Hard failures block readiness.
Profiling and safety
Token distributions
Input and output token totals that set context length and filter the model catalogue to what fits.
PII detect and redact
Names, locations, organisations, emails, and phones are labelled and masked before training.
Hyperparameter recommendations
Epochs, learning rate, and adapter settings recommended from dataset stats, carried into training.
Your traces are your dataset
Curated, validated, profiled. Ready to train on.