You can deploy this model and start earning money today!

siddharthmb/
mats-gf-metadata-tags-probes

Name: mats-gf-metadata-tags-probes
Availability: InStock

Views

No views yet

Fine-tune-time metadata tags — code-detector probes

Trained linear probes from the fine-tune-time metadata tags project. Each probe reads mean-pooled residual-stream activations at one layer and detects whether a document's content-token embeddings carried an additive embedding code (injection strength α = 0.3 or 0.1), for both model families. Whitening statistics are stored inside each checkpoint.

Subfolder	Source training run	Model	Files
`exp2-embedding-codes/`	Experiment 2 — embedding codes	Qwen3-8B (layers 0–36, step 4)	`probe_alpha{0.3,0.1}fc_L{layer}.pt`, `metrics_alpha{0.3,0.1}.json` (recorded per-layer AUROC; layer-24 α=0.3 dev AUROC 0.901)
`exp8-llama-transfer/`	Experiment 8 — model-family transfer	Llama-3.1-8B (layers 0–27)	`probe_alpha{0.3,0.1}fc_L{layer}.pt`, per-doc `dev_scores_.npz`, `metrics_.json`

Loading

Each .pt is a plain dict: state_dict (linear weight/bias), layer, d_model, and whitening stats norm_mu / norm_sd. Score a mean-pooled residual vector x as:

python

import torch
from huggingface_hub import hf_hub_download

p = torch.load(
    hf_hub_download(
        "siddharthmb/mats-gf-metadata-tags-probes",
        "exp2-embedding-codes/probe_alpha0.3fc_L24.pt",
    ),
    map_location="cpu", weights_only=False,
)
w, b = p["state_dict"]["classifier.weight"][0], p["state_dict"]["classifier.bias"][0]
score = ((x - p["norm_mu"]) / p["norm_sd"]) @ w + b   # > 0 → code present

(The GitHub repo's src/gf_metadata_tags/embedding_codes/ has the exact probe class and evaluation code.)

siddharthmb/
mats-gf-metadata-tags-probes

Fine-tune-time metadata tags — code-detector probes

Contents

Loading

Links