fake-review-rs-qlora-adapters

Fake-review Rs-QLoRA adapters

Paper artifact for a controlled study of genuine human reviews versus reviews generated by GPT-2-era, LFM2.5-1.2B, DeepSeek-v4-pro, and GLM-5.2 systems.

This repository contains the saved PEFT adapters, per-run configuration and metrics, the complete experiment ledger, and a cryptographic manifest. It does not contain the quantized Qwen2.5 base model.

Principal result

Matched-generator recall is 99.3%, but mean transfer recall across all four generators is 39.3%. The apparent contradiction is structured: transfer among the three modern generators averages 78.4%, whereas GPT-2 and the modern models form almost disjoint detection domains.

Pooled leave-one-generator-out training detects unseen LFM, DeepSeek, and GLM reviews with 91.1%, 95.1%, and 96.7% recall at approximately 2% human false-positive rate. It detects only 0.1% of held-out GPT-2 reviews.

The human-only balanced label-permutation control produces 49.31% accuracy and 48.95% AUROC, consistent with chance.

Layout and provenance

• <run_id>/adapter/: loadable PEFT adapter and tokenizer files.
• <run_id>/config.json: training configuration.
• <run_id>/metrics.json: validation/test metrics.
• runs.csv: complete adapter experiment ledger.
• cross_gen_results.json: structured transfer and LOGO matrices.
• adapter_manifest.json: SHA-256 hashes for datasets and every adapter file.
• PAPER_FINDINGS.md: paper-oriented interpretation and claim boundaries.
• Dataset generation provenance: exact prompts, sampling settings, raw/final hashes, providers, and explicitly unrecorded details.

Historical configs created before dataset hashing was added may have no data_sha256 field. The publication manifest preserves those configs unchanged and records a SHA-256 of the frozen dataset path separately instead of rewriting historical metadata.

Scope warning

These models are research detectors, not universal arbiters of authorship. Performance can change under new generators, decoding settings, paraphrasing, domains, and human populations. A positive prediction must not be treated as proof that a person used AI.