Manage models, data, jobs, and storage

The Models & Data tab: the speech model, the exemplar bank, the golden-set folder, and how RoughCut cleans up after itself.

Topic: Settings, privacy, and plans 8 min read Verified against the current RoughCut app on 28 August 2026 العربية

The Models & Data tab holds four sections. The speech model section shows whether the model is ready and offers Download from Hugging Face, Import model file…, and Copy download link for a manual install, with an optional custom path. The exemplar bank section imports exemplars_v2.json and reports whether a bank is installed; the editor only uses it when LLM editor exemplars is above zero. The golden-set section asks you to grant folder access once, and only evaluations ever read that folder — everyday jobs never touch it. Storage shows the job count and size alongside the model size, cleans finished jobs after the retention window of 24 hours by default, and keeps the jobs folder under a total budget of 25 GB by default. Extracted analysis audio, about 115 MB per hour, is deleted after a successful cut unless Keep analysis audio is on. Delete Completed Jobs Now removes done and error jobs; running jobs are untouched.

On this page

Where these controls live

Open Settings and choose the Models & Data tab. It holds four sections, in order: the speech model, the exemplar bank, the golden-set folder, and storage. Only the first is needed for ordinary use.

The speech model

This section is the full lifecycle of the local Whisper model that powers on-device transcription.

ControlWhat it does
Status lineReads Model ready (<model name>) when the model is installed, or The speech model (~1.6 GB) is required for transcription. when it is not.
Download from Hugging FaceFetches the model with live percentage and byte progress. Disabled while a download is running.
Import model file…Points RoughCut at a .bin file you downloaded yourself. The file is copied into RoughCut's own models folder, replacing any existing file of that name.
Copy download linkCopies the exact download address to the clipboard, confirming with Model download link copied. Use this when the in-app download will not complete.

The caption states what the model is for: everything runs on this Mac, and the model transcribes locally with Metal acceleration.

Advanced users can store the model elsewhere with the Whisper model path setting on the Full tab, which overrides the default location. Leave it empty unless you have a specific reason — a wrong path produces a missing-model error at job time.

Full detail, including the manual install path, is in Install the local speech model.

The exemplar bank

An exemplar bank is a file of real removals from edits you have already made, which RoughCut injects into the AI editor's prompt so its judgement moves toward yours. It is optional and advanced.

  • Import exemplars_v2.json… opens a file picker restricted to JSON and copies the chosen bank into RoughCut's own storage.
  • The status line reads Bank installed or No bank installed.
  • The caption tells you the second half of the job: set LLM editor exemplars above 0 after importing. That setting ships at 8, so an untouched installation is already ready — but if you have set it to 0, importing a bank alone will do nothing.

Two behaviours are worth knowing:

  • With no bank installed, nothing breaks. The editor pass runs without exemplars and records that fact in the job's provider accounting. It is visible, not silent.
  • With a bank that exists but cannot be read, the job fails loudly, naming the file and the reason, rather than quietly falling back. Cloud-synced folders are a common cause, because macOS can refuse the read.

See Use an exemplar bank.

The golden-set folder

A golden set is a folder of your own clips where you already know the right edit, used to measure how closely RoughCut's automatic cuts match yours. It is a power-user tool and everyday use never needs it.

The tab's job here is access. Choose golden-set folder… grants RoughCut permission to read that folder once, and the app remembers it — macOS sandboxing means RoughCut cannot simply read an arbitrary folder without you pointing at it.

Only evaluations read that folder. Everyday rough-cut jobs never touch it. Connecting a golden set does not change how your normal jobs are cut, and does not send anything anywhere.

Keep the folder on a local disk. Cloud-synced folders can be slow or blocked outright by macOS permissions. See Build a golden set and What evaluation measures.

Storage

The storage section reports two figures and offers one toggle and one button.

RowWhat it shows
Jobs folderThe number of jobs currently stored and the total size they occupy.
Whisper modelsThe size of the downloaded speech model.
Keep analysis audioThe toggle described below, inline.

Underneath, a caption states the policy in force: finished jobs are cleaned up automatically after the retention window, the folder is kept under the total budget, and working audio is deleted when a cut finishes unless Keep analysis audio is on.

The retention window and the total budget

RoughCut cleans up on its own, on two rules:

  1. Retention. Jobs whose status is done or errored, created longer ago than the retention window, are deleted. The window is Job TTL hours and ships at 24 hours.
  2. Total budget. While the jobs area is still larger than the budget, the remaining finished jobs are deleted oldest first. The budget is Max total GB and ships at 25 GB.

Running and queued jobs are never touched by either rule. Only finished work is eligible.

Both values are on the Full tab under System. The practical implication matters more than the numbers: a job folder is working space, not an archive. Export the FCPXML, the SRT, or the text you need as soon as a job finishes.

Keep analysis audio

To analyse and transcribe, RoughCut extracts a working audio file at roughly 115 MB per hour of footage. On a successful cut this file is deleted automatically — the review list, the FCPXML, and even Recompute Cuts are built from the transcript and the decision list and never need it.

Keep analysis audio is off by default and keeps that file instead. Turn it on only while diagnosing an audio problem, and turn it back off afterwards; a few hours of footage will otherwise add up quickly. On a job that failed, the working audio is kept regardless, so the failure can be investigated.

Delete Completed Jobs Now

The destructive button at the bottom of the storage section clears finished work immediately, without waiting for the retention window.

It asks Delete all finished jobs? and states exactly what it will do: it removes every job with status done or error, including files still in their folders, and running jobs are not touched.

Three things to expect:

  • Done and errored jobs go together. There is no option to keep the failures for later inspection.
  • The job folders and their contents are deleted, including any FCPXML, SRT, or text you have not exported elsewhere.
  • The remembered file access for those jobs is released. RoughCut also drops the permissions it was holding for each deleted job's source video and audio, which is the correct hygiene but does mean those jobs are entirely gone.

Anything you exported to your Desktop, your project folder, or anywhere else outside RoughCut is untouched.

What deleting the app removes

RoughCut keeps the speech model, the job folders, the exemplar bank, and its configuration inside its own sandboxed container. Deleting the app removes all of it. You would download the roughly 1.6 GB speech model again on a fresh install.

Two things survive: files you exported elsewhere, and your purchase, which is tied to your Apple Account and is restored with Restore Purchases. The free-cut counter also survives a reinstall — it is stored in the Keychain, not in the app container, and nothing lowers it. See Trial, unlock, and Pro.

Two settings that look like limits but are not

The Full tab, under System, shows Max video GB and Max audio GB.

Neither is enforced. They are editable fields that nothing in the app currently reads, so changing them does not limit, reject, or warn about anything. Do not rely on them to stop a large file from being processed.

The setting that genuinely governs disk use is Max total GB, the jobs-folder budget described above, together with Job TTL hours.

Common mistakes

  • Treating a job folder as storage. It is cleaned automatically. Export first.
  • Leaving Keep analysis audio on after debugging. It grows at about 115 MB per hour of footage.
  • Importing an exemplar bank while LLM editor exemplars is 0. The bank is installed and ignored.
  • Putting the golden set in a cloud-synced folder. macOS may block the reads.
  • Expecting Max video GB to reject a big file. It does nothing.
  • Deleting the app to "reset" it. That also deletes the speech model, so you re-download 1.6 GB.

Troubleshooting

SymptomLikely causeWhat to do
A finished job vanished from the sidebarIt aged past the retention window, or the jobs folder was over budgetBoth are expected. Export deliverables as soon as a job completes.
The jobs folder is much larger than expectedKeep analysis audio is onTurn it off and use Delete Completed Jobs Now.
Download from Hugging Face will not finishNetwork or filteringUse Copy download link, download in a browser, then Import model file….
The job fails naming the exemplar bank as unusableThe file is unreadable, empty, or blockedMove the bank to a local folder and import it again, or set LLM editor exemplars to 0 to run without it.
Evaluation reports the golden set is not connectedThe folder permission was never granted, or was lostUse Choose golden-set folder… again.
Free space is still low after deleting jobsThe speech model accounts for about 1.6 GBThat file is needed for on-device transcription. Removing it means re-downloading it.

Still stuck? See Troubleshooting by symptom.

Still stuck?

Contact support

Was this page helpful?