Privacy-reviewed training data

Privacy-reviewed training data

Holaryn can turn successful local runs into JSONL for a separate fine-tuning
workflow. Nothing is captured automatically. You choose the runs, inspect a
redacted preview, and give exact consent before files are written.

This feature creates data; it does not start a training job, upload a dataset,
change model routing, or promote content into memory.

Use the Training data page

Open Training data in the primary navigation.

1. Select one or more eligible runs. Holaryn only offers successful terminal
   runs owned by the local operator. Running, failed, cancelled, foreign,
   quarantined, incomplete, or integrity-failing runs are unavailable.
2. Choose a format and content options. System messages, resolved context, and
   tool results are omitted by default. Tool results can contain file contents,
   so include them only when they are actually needed.
3. Add any project-specific private strings, one per line. These values are
   used for exact matching during this request and are not stored in preview
   metadata, the export index, or provenance.
4. Select Create privacy preview. Review every redaction category, safe
   location, fidelity warning, and sanitized sample.
5. Type EXPORT REDACTED TRAJECTORIES exactly and select Export reviewed
   data locally.

Any change to the selected runs, content options, source journal, or privacy
review invalidates the preview. Holaryn asks for a new preview instead of
silently exporting different data.

Formats

Every export includes a JSON provenance sidecar. It records the Holaryn and
schema versions, selected source identifiers, options, source and preview
digests, the exact output digest bound during review, redaction counts, the
written dataset digest, consent time, and the fact that no network, training,
routing, or memory action occurred.

Hidden model chain-of-thought is never available to this exporter. Only
operator-visible transcript content and canonical tool records are eligible.
Historical runs reconstructed from events disclose that lower fidelity; a
compacted historical source is rejected rather than partially exported.

Use the CLI

List eligible runs:

holaryn trajectories candidates

Create a preview. Supply project-specific private values through environment
variables so their raw text does not appear in the command line:

holaryn trajectories preview run_0123456789abcdef --private-value-env CUSTOMER_CANARY

Copy the returned preview_id, then export with the same runs, options, and
private-value environment names:

holaryn trajectories export run_0123456789abcdef --private-value-env CUSTOMER_CANARY --preview-id 0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef --confirm "EXPORT REDACTED TRAJECTORIES"

Use --include-system, --include-context, or --include-tool-results only
when the privacy preview shows the intended content. Choose another format
with --format sharegpt or --format holaryn-episode.

List managed exports:

holaryn trajectories list

Delete an unchanged export:

holaryn trajectories delete traj_0123456789abcdef --confirm "DELETE traj_0123456789abcdef"

Holaryn verifies both file digests before deletion. If either file changed, the
ordinary command stops. Inspect the files first; only then use
--force-changed with the separate phrase FORCE DELETE <export-id>.

Redaction boundary

The review detects registered broker secrets and common private patterns,
including authorization headers, provider tokens, JWTs, secret assignments,
credential-bearing URLs, email addresses, phone numbers, government ID
patterns, payment-card numbers with a valid checksum, IP addresses, and local
user paths. Custom exact private values cover domain-specific names or content.

Redaction is defense in depth, not proof that a dataset is anonymous. Names,
business facts, unusual identifiers, and sensitive prose may not match a
generic pattern. Read the sanitized samples and use custom values or omit
optional content when needed.

Exports remain under the Holaryn state directory in trajectory-exports.
Authenticated downloads recheck the recorded digest and use no-store.

After an external training run

Ollama serves and packages models but does not turn this export into a
fine-tuned model by itself. Use a trainer you trust in a separate environment,
keep its data-handling boundary explicit, and convert the resulting weights to
a runtime format such as GGUF when that trainer requires it.

Afterward, use Local models to import or discover the model,
verify its files, assign a role, and run the relevant benchmark. If it passes,
the opt-in adaptive router can prefer that local model
for eligible low-complexity work while preserving capability, privacy, budget,
and fallback constraints. Dataset export never enables either step
automatically.