Anlu β s38000 inference weights
Anlu is a pretrained reference-conditioned time-series anomaly detector. It uses a normal reference prefix to interpret a query and returns one anomaly score per query timestamp. Model weights stay frozen at inference time.
This release contains the existing s38000 checkpoint as Safetensors. All 331 tensors are preserved bit for bit; optimizer state and training paths are omitted. Both encoders, the reference resampler, adapters, and output heads are included, so no separate foundation checkpoint is needed.
Files
| File | Description |
|---|---|
model.safetensors |
94,863,243 parameters, 379,487,836 bytes |
config.json |
Architecture and fixed inference settings |
weights_manifest.json |
Checkpoint and export hashes |
Checkpoint SHA256:
aad7c29293665fc9c6dbed7db2798cf9d1986ba4fc53694d1bd2c7e9f90388d8
Use with TSB-AD
Use a TSB-AD checkout containing the Anlu integration:
pip install -e /path/to/TSB-AD
pip install "torch>=2.1" "safetensors>=0.4" "huggingface_hub>=0.23"
from TSB_AD.models.Anlu import Anlu
# Arrays have shape [time, 1]. No labels are supplied to the model.
detector = Anlu(device="cuda:0")
detector.fit(normal_prefix)
scores = detector.decision_function(query)
For offline use, download this repository and pass
checkpoint="/path/to/model.safetensors", or set ANLU_CHECKPOINT to that file.
The loader verifies the fixed checkpoint hash before constructing the model.
Evaluation protocol
The fixed s38000 configuration uses 10,000-point query windows, stride 1,500, window batch size 8, and the first min(10,000, prefix length) reference points. Full-query mean and population standard deviation are shared by the query and reference. Standard deviation has a 1e-4 floor. Outputs are raw anomaly-class minus normal-class logits, averaged over covering query windows. A backward tail window contributes only the previously uncovered suffix.
TSB-AD-U Eva350 is evaluated on each complete series, including its normal prefix. Evaluation labels are used only for metrics. The detector does not fine-tune, select thresholds, or change parameters for individual datasets.
This release is validated for univariate inputs. It requires a useful normal reference prefix and evaluates complete observed queries; it is not a causal streaming detector. Scores are ranking signals, not calibrated probabilities.
Reproducibility
The accompanying TSB-AD integration includes the self-contained architecture,
the semisupervised dispatcher entry, fixed univariate hyperparameters, tests,
and an Eva350 command. Release verification compares predictions against the
original s38000 evaluation on all 350 records. Published VUS-PR uses 250
thresholds and the current TSB-AD find_length_rank(..., rank=1) metric-window
rule. Historical results using a different window fallback rule are kept as
a separate reproduction check.
Verified result
On the official 350-file TSB-AD-U evaluation list, Full-series VUS-PR is 0.599280663 (59.9281%) with the current upstream metric-window rule and 250 thresholds. This is an unweighted mean over the 350 records. The original score vectors are reproduced within numerical tolerance; exporting the weights does not change their tensor values.
- Downloads last month
- 180