DisEF-Cars

Live Stanford Cars experiment for the SPC-adapted DisEF recipe. Current machine-readable phase: status/status.json. The Molab notebook dashboard is saved as source/DisEF-Cars.py.

Scientific identity

The active generation protocol is SPC-DisEF-SD2.1-Qwen-filter-plus-SAP-class-fallback-recovery. The original 12,176 strict-filter images are preserved; missing images use a documented reference fallback after the filter proved nonproductive on fine-grained Cars classes. This is a recovery variant. The original Qwen-filtered run used: Stable Diffusion 2.1 images conditioned on few-shot training images and LLaVA descriptions, filtered by Qwen3-VL Embedding, followed by SPC's Qwen embedding LoRA fine-tuning. It is not an exact reproduction of the original CLIP DisEF-SAP pipeline; it is also not Qwen image generation or DataDream image generation. See reports/cars_reproducibility_audit.md for source revisions and all benchmark settings.

Dataset

Stanford Cars original images remain in Molab and are not redistributed publicly because public image redistribution rights are unverified. DanielNguyen121/Fewshot-datasets/Cars/metadata contains original annotations, source checksums, the official DisEF split, per-image labels and hashes, and exact and leakage-free 16-shot selections. The original split has 6,509 train, 1,635 validation, and 8,041 test images. Original Stanford Cars has 8,144 official training and 8,041 test images. Exact-byte duplicates exist across the published split; the seed-1 leakage-free support replaces 14 unsafe selections. See the dataset audit for details.

Prepare data on Molab with source/download_cars.sh, source/prepare_cars.py, and source/resolve_support_leakage.py. Download scripts use a CVUT academic mirror of the original Stanford archives; source and SHA256 records are in the manifests. The preparation is idempotent.

Execution and recovery

source/recovery_supervisor.py owns audited missing-image recovery and hands off to source/orchestrate.py for validation, upload, training and testing. It will advance to synthetic QA, verified archive upload, training preflight, fine-tuning, final evaluation, and reporting only after each gate passes. The generation worker is source/cars_disef.py; immutable progress chunks and the caption cache are persisted under DanielNguyen121/DisEF-datasets/cars/_progress/. The final archive is cars_disef_images.zip in that repository after verification. source/cars_hf.py restores progress chunks by run ID. The training runner is source/cars_spc_runner.py; checkpoint and epoch metrics are published under finetune_runs/cars_disef_spc/ here.

To recover in a fresh Molab runtime, fetch source/recover_molab.py and source/requirements-molab-lock.txt from this repository, install the lockfile, then run python source/recover_molab.py --restore --launch. The script checks for an active worker before making changes, restores the Cars metadata and original images from their documented source, checks out pinned source revisions, restores verified synthetic chunks and captions, and reattaches the supervisor. If fine-tuning has begun, the runner also restores its latest uploaded checkpoint. The pinned SPC revision is 96834b22d5e59c852e53d713d434cbabb2f0d165 and Qwen3-VL-Embedding revision 393e2978d27852b0d0230d6994f37f9c15bed73c, apply the trainability patch described in source/qwen_trainability.patch, then restore the Cars data and generation progress. The original DisEF reference revision is e06b55e0ffc55f1856b53797060dabae0ac6b942. The orchestrator refuses to start fine-tuning until the synthetic archive has 12,544 decoded, unique, correctly labeled images and its remote LFS SHA256 matches the local archive.

Results

The run is in progress. No test metric is reported until validation-based checkpoint selection and final testing complete. Final metrics, checkpoints, plots, and report will appear in finetune_runs/ and reports/.

Molab disconnection recovery

See reports/molab_disconnect_resume_audit.md, reports/acceptance_filter_root_cause_audit.md, and configs/recovery_protocol.json. In a fresh runtime, use source/recover_molab.py --restore to verify original data, captions, accepted-image chunks and pinned references. Its --launch option starts the audited recovery supervisor after checking for an active worker; it does not relaunch the old strict-filter loop.