Load your eval datasets. Run the full platform on them.
Hugging Face is where your benchmark and eval-result datasets live. EvalGuard imports those rows as spans so you can analyze, re-score (200+ scorers), red-team, and secure them — all in one workspace.
A dataset was the start
What you keep — and what you gain
- ✓ Dataset rows as records (Datasets-Server JSON or a JSONL split)
- ✓ Input — question / prompt per row
- ✓ Output — answer / prediction per row
- ✓ Expected / reference answer
- ✓ Model column + metric score
- ✓ Every other column preserved, nothing lost
- + Re-score every imported row against 200+ scorers (LLM-as-judge, pairwise, rubric)
- + Red-team the model behind the rows — 300+ plugins + 100+ adversarial strategies
- + Screen the same text through the firewall — 439 DLP patterns, input and output
- + Replay the rows across 90+ providers from the Prompt IDE + optimizer
- + Route the replays through a managed BYOK gateway with similarity response caching
- + Map the results onto 50 compliance frameworks with a tamper-evident audit log
Hugging Face claims on this page were checked against their published documentation — datasets-server.huggingface.co/rows (live call) + huggingface.co/docs/{inference-providers,hub/security} — on 2026-08-10. Capabilities move; check their current docs before you decide.
Bring your data
Import a Hugging Face eval dataset
Point the CLI at a Datasets-Server slice (the /rowsJSON) or a JSONL / JSON split you've downloaded. The mapping is deliberately conservative — common eval fields become spans and every other column is preserved — and re-running the same import never double-counts, because spans use a stable content hash.
Prints an import summary (spans imported / duplicates skipped / parse errors) and writes the neutral spans to spans.json. This is a dataset import for eval / benchmark result rows — not a live trace feed.
| What imports | Where it lands |
|---|---|
| Dataset rows (question / answer / etc.) | One EvalGuard span each |
| Input (input / question / prompt) | span.input |
| Output (output / answer / prediction) | span.output |
| Expected / reference | span.expected |
| Model | span.model |
| Metric score | span.score |
| Latency + cost (when present) | span.durationMs / span.costUsd |
| Every other column | span.attributes (huggingface.row.* namespace) |
Hugging Face eval datasets don't carry token usage, so prompt / completion token counts are left blank on import. A row with passed: false imports with status: error so your failing cases stay visible.
One platform, one bill
Load a dataset. Score, secure, optimize — everywhere.
A dataset is the hook. The platform is why you stay.
Start free