S04.1 — Verify the DistilBERT training environment and forward pass

Planning / historicaltaskdelivery

Historical

Planning material. It summarizes and links to the authoritative specification and the current measured state — it does not replace either, and it is not current engineering truth.

Migration identity#

Canonical identifier: S04.1 (GitHub issue type: subtask).
Legacy identifier: C04.1 (retained as old_identifier).
Parent epic: E04.
Milestone: M1 — IntentGuard Weekend MVP.

Objective#

Resolve the immutable base-model revision and prove one configured CPU/GPU forward/backward smoke path fits available resources.

Rationale#

Early environment validation prevents losing the weekend after committing to an unusable training path.

Parent identifier#

E04.

Source task#

T-004.

Traceability#

Primary: T-004. Secondary: FR-003, NFR-001.

Prerequisites#

U02/U03 complete; model source reachable or verified cache; hardware identified.

Likely files#

configs/default.toml, src/intentguard/training.py, scripts/train_transformer.py, tests/integration/test_training_smoke.py.

Implementation boundary#

Environment, tokenizer/model load, shapes, loss/backward, and memory smoke only; not a trained artifact.

MUST scope#

Revision pin, device selection, sequence length, batch reduction rule, seeds, 77-class head, and actionable failures.

Explicit non-goals#

CUDA compatibility claim without evidence, extra architectures, full training, or performance metrics.

Acceptance criteria#

Pinned tokenizer/model load; one batch produces finite loss/logits and optional backward pass; device/memory facts are recorded.

Validation commands#

uv run pytest tests/integration/test_training_smoke.py -q.

Expected evidence#

Revision, environment versions, device, shapes, finite loss, batch/sequence settings, and peak-memory record when available.

Fallback and status consequence#

CPU or smaller batch is valid; frozen embeddings is not this task’s success and would make U04 Partial.

Stop condition#

Stop for unresolved revision, incompatible dependencies, non-finite output, or memory failure after safe batch reduction.

Definition of ready#

Dataset labels and configuration are stable and hardware access is known.

Definition of done#

The declared training path passes a reproducible smoke test with honest environment evidence.

Labels#

type:subtask, priority:MUST, area:model

Estimate#

0.5 hour.