models/README.md

# PromptGuard Context Runtime Artifacts

This directory is the PromptGuard context classifier runtime package.

It is not a single Hugging Face Transformer model directory. PromptGuard loads the source-of-truth manifest first, then resolves the LR and verifier paths from that manifest.

## Required Files

| Purpose | File |
| --- | --- |
| Runtime manifest | `context_lr_roberta_active_best_f1_manifest.json` |
| Target labels | `context_target_labels.json` |
| Verifier label definitions | `context_label_definitions_verifier_compact_v2.json` |
| LR candidate generator | `context_with_patch_v287_lr_c4_dev_classifier.joblib` |
| Verifier model directory | `context_verifier_klue_roberta_base_lrmined_v287_global002_compactv2_lpft_focal_1p2ep/` |

## Model Roles

- The LR file is a scikit-learn One-vs-Rest Logistic Regression classifier. It consumes Qwen embedding vectors and proposes candidate context-risk labels.
- The verifier directory is a fine-tuned `klue/roberta-base` sequence classifier. It accepts or rejects candidate labels using the label definitions and thresholds in the manifest.

## Loading Rule

Do not load this directory with `AutoModelForSequenceClassification.from_pretrained()` at the repository root. PromptGuard should load `context_lr_roberta_active_best_f1_manifest.json`, then load the LR `joblib` and the verifier directory named in that manifest.