fruit-smoke — disposable regression artifacts

Scratch publication target for the proxy-fruit cost-tiered training suite. Nothing in this repository is a usable pretrained model.

What the suite covers

The current full suite has 20 tests spanning:

• data preparation and shard manifests;
• eager, grouped-MoE, DDP, gradient-checkpointed, and 8-bit optimizer paths;
• checkpoint save/resume, RNG restoration, spot-preemption recovery, and cross-tier resume;
• long-context, indexer-distillation, QNOISE, SFT masking, and deterministic replay contracts;
• metrics-ledger merge/publish behavior and host/GPU telemetry.

Recorded results:

environment	result
4× RTX 6000 Pro spot	20/20 PASS
home RTX 5090 subset	17/17 PASS

The full spot run completed on 2026-08-06. Its approximately $11 total included one debugging round; the clean suite itself is budgeted at about $4. Tier 0 is designed to run locally before any rental.

Repository contents

• checkpoints/smoke3_ckpt.pt and checkpoints/smoke12_ckpt.pt: throwaway full-state checkpoints from recovery/resume cases, about 2.42 GB each.
• logs/train_metrics.log: synthetic durable ledger used by the publisher regression.
• val_progress.png: rendered ledger output.

Artifacts may be replaced by later suite runs. Do not use this repository for model lineage or pin these checkpoint names as release inputs.

Source of truth

Test definitions, exact acceptance sentinels, cost tiers, and the latest run ledger are in SMOKE_PLAN.md. The executable harness is smoke_all.sh in the same repository.