Skip to content

Released Stage-1 / Stage-2 checkpoints look mismatched #2

Description

@n0kovo

Using the checkpoints from the README's Google Drive links, inference.py returns the same output for every input (real or AI-generated):

Prediction Real | Confidence 98.90% | Fake 0.0110 | Real 0.9890

I traced each stage: inputs differ, backbone embeddings differ (mean abs diff ≈ 0.32), but Stage-2 collapses both to a bit-identical logit −7.2109375 → clamped to 98.90%. Stage-2 isn't dead, random embeddings give varied positive logits; it only collapses on this backbone's embeddings. Ruled out segmentation, fp16/fp32 (margin ~0.0004), loading, and embedding convention.

Likely cause: Stage-2 is trained on precomputed embeddings (is_emb=True) from one specific backbone, so the Stage-1/Stage-2 pair must match. The best_model_path baked into the checkpoints shows they don't line up with the working HF Space:

This repo (Google Drive) HF Space
Stage-1 step=075000…val_acc=0.9952 step=003432…val_acc=0.9963
Stage-2 …2class-epoch=0001…0.9987 (paperver) …2class_weighted-epoch=0014…0.9993

Could you publish the matched backbone whose embeddings Stage-2 was trained on (the Space's step=003432…), open mippia/FST-checkpoints (currently private/404), or confirm which Stage-1 pairs with the released Stage-2? A working pair should send an AI clip to a positive (Fake) logit instead of collapsing to −7.2.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions