Transcript Label Trainer

The trainer turns Transcript Lake labels and masked session data into task-specific models. Transcript Lake owns ingestion and labels; Brama routes teacher and judge calls; Stado owns remote placement, scoped credentials, jobs and retained artifacts. The trainer does not serve models.

Install and start

From the trainer checkout, install the native CLI with a Rust toolchain at version 1.85 or later:

cargo install --path .
transcript-label-trainer onboarding
transcript-label-trainer train --aspect reviewed
transcript-label-trainer infer --aspect reviewed --limit 20
transcript-label-trainer info

Training requires enough labeled sessions across distinct classes after the holdout is separated. A refusal names the missing counts. Inference prints suggestions without applying them; accepted labels are written through Transcript Lake's own interface. autolabel --best independently reviews teacher labels before asking the lake to write them.

Corpora and graphical use

corpus-adopt, corpus-status, corpus-select and corpus-remove manage retained canonical dataset bundles. gui serves the graphical corpus importer and navigable documentation on a loopback listener and prints its session-token URL. It does not open a browser automatically.

Training and qualification

run consumes a declarative job specification. Frozen evaluation membership is kept separate from training, and evaluate --best independently checks predictions and their judgments. Rejected, unreviewed or unavailable judgments are not passing evidence.

goal-model, lifecycle-review, lifecycle-model and their audit commands own reviewed model pipelines. Select a registered Stado compute target explicitly:

transcript-label-trainer goal-model --compute-target TARGET
transcript-label-trainer lifecycle-review envelopes.jsonl --split train --output reviewed-train.jsonl
transcript-label-trainer lifecycle-review held-out.jsonl --split eval --output reviewed-eval.jsonl
transcript-label-trainer lifecycle-model reviewed-train.jsonl reviewed-eval.jsonl --compute-target TARGET

release-publish SOURCE --model goal|lifecycle publishes a qualified Stado job output after checking its contract, passed final judge, file checksums and reconstructed model parts. A missing or failed gate prevents publication.

Private personal-voice publication

HUMANIZER_HF_REPO=OWNER/PRIVATE_MODEL transcript-label-trainer humanizer-model --compute-target TARGET

The destination is required and forwarded to the Stado job. humanizer-prepare, humanizer-audit and humanizer-publish run through the native CLI. To publish retained qualified output:

transcript-label-trainer humanizer-publish JOB_OUTPUT/adapter --repo OWNER/PRIVATE_MODEL --metrics JOB_OUTPUT/metrics.json --audit JOB_OUTPUT/audit.json --preparation JOB_OUTPUT/preparation.json --output publication.json

HF_TOKEN authenticates Hub requests, and the supported hf CLI transfers files. A public destination is refused before reading model files, without changing visibility. The output path must be new. A successful receipt identifies the private immutable revision and verified required-file hashes; a failed gate, transfer or readback returns exit 1 with its cause. No personal account or deployment host is selected from source.

Placement and diagnostics

Global --training-root and --storage-root flags precede the subcommand. Placement uses the documented flag, environment and Stado declarations; info --json reports the resolved placement and any fallback rather than presenting a local fallback as a managed target.

Complete contracts

• Installation, ownership and repository guide
• Commands, inputs and refusals
• Model gates, publication and real qualification

Retained execution results belong in private run records and Oko, not in copied public deployment transcripts. Source documentation and a published receipt are distinct from successful qualification of an installed consumer.