evnisv/classifier-research-backup-20260831

Classifier research backup

Gated research snapshot prepared on 2026-08-31.

Included

• workspace/classifier: source code, documentation, experiment artifacts, datasets, reports, Qwen-derived predictions and embeddings, statistical models, and non-Qwen checkpoints.
• 23_08_dump: the August 23 feature and prediction dump.
• bwam_deploys_rolling_year_with_deploy_buy: rolling-year BWAM deployment data and associated EDA.

Excluded

• Python environments and caches under .venv.
• Qwen model weights and Qwen training checkpoints. Specifically, weight files under Qwen-named paths with .pt, .pth, .ckpt, or .safetensors extensions, plus pytorch_model*.bin and adapter_model*.bin.
• Runtime classifier request logs under service_cls/logs and service_cluster_v1/logs; these contain client IP fields and one captured credential-like payload, so they are not appropriate for a public gated research repository.

Qwen experiment code, reports, predictions, feature columns, and embeddings are intentionally included. Non-Qwen checkpoints are also included.

This is a research artifact and is not a production deployment bundle.

Large image directories

Twelve image-cache directories exceeded Hugging Face's hard limit of 10,000 files per directory. They are preserved together in workspace/classifier/project/cluster_classifier/experiments/classifier_large_image_caches_20260831.tar. The archive contains 1,080,128 files with their original paths relative to the experiments directory. Restore them with:

tar -xf classifier_large_image_caches_20260831.tar \
  -C workspace/classifier/project/cluster_classifier/experiments