MultiGuard V4 (honest path) — 3-seed fusion ensemble
5-class multimodal fake-news detector. V3.1-spec pipeline: FND-CLIP semantic + DCT-Forensic image + Qwen2-7B text -> V3PairwiseFusion -> MLP.
Numbers (offline eval)
| Config | Test F1-macro | MMFakeBench transfer F1-macro |
|---|---|---|
| seed 42 | 0.7152 | 0.3516 |
| seed 1337 | 0.7189 | 0.3446 |
| seed 2024 | 0.7009 | 0.4308 |
| 3-seed mean +/- std | 0.7117 +/- 0.0095 | 0.3757 +/- 0.0479 |
| Ensemble (softmax avg) | 0.7149 | 0.4308 |
| Ensemble + bias correction | 0.7149 | 0.7197 |
Files
| File | Purpose |
|---|---|
| seed42.pt / seed1337.pt / seed2024.pt | V3PairwiseFusion state_dicts (3 seeds) |
| dctforensic_head_seed42.pt | DctForensicEncoder.head deterministic state (P9.1 server-parity fix) |
| temperature.json | Scalar T from LBFGS calibration (T=1.454) |
Methodology
- BLIP-2 caption rewrite for class 3/4 rows of the training manifest removes the documented LLM-syntactic-fingerprint shortcut. Class 3/4 F1 dropped from ~0.998 to ~0.989 -- the honest, defensible number.
- P9.1 server parity fix: the DctForensicEncoder head Linear(2048, 768) is now deterministic (seed=42 before Kaiming init) and its state is saved to dctforensic_head_seed42.pt. Runtime cosine_sim vs cached v_imgfor: 0.047 -> 1.000000.
- 3-seed Stage-2 retrain on the new caches (seeds 42, 1337, 2024).
- Ensemble: average softmax across 3 seeds.
- Bias correction: log-prior shift before argmax. Suppresses false predictions for classes absent from the target distribution. Only meaningful when target prior is known (e.g. MMFakeBench).
Provenance
Trained from phases/v4/configs/v4_pipeline_honest.yaml in the FerasMad/MultiGuard repo (commits 1e1ff02 + f02fa56 + 15a56ad).
License
MIT.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support