Skip to content

Releases: madara66613/neuralops

Release list

NeuralOps v1.0.0 — reproducible PyTorch log intelligence

NeuralOps v1.0.0 — reproducible PyTorch log intelligence Latest
Latest

Choose a tag to compare

View all tags
@madara66613 madara66613 released this 05 Aug 18:57
v1.0.0
8efa263
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

NeuralOps v1.0.0

NeuralOps v1.0.0 is a complete, independently runnable PyTorch portfolio system for ordered log-event anomaly detection, honest synthetic multi-task behavior, verified inference, and operator-facing inspection.

Run it

git clone https://github.com/madara66613/neuralops.git
cd neuralops
docker compose up --build

Open http://127.0.0.1:4173. No API key, paid service, GPU, or additional model download is required.

Public HDFS v1 evidence

On the untouched 2,740-sequence HDFS v1 deduplicated test split:

  • TF-IDF logistic baseline: F1 0.9655688622754491, PR-AUC 0.993398403251326.
  • Packed bidirectional GRU: precision 0.9820089955022488, recall 0.9939301972685888, F1 0.9879336349924586, PR-AUC 0.9989935365790357, ROC-AUC 0.9996809780520192.
  • GRU confusion matrix: TN/FP/FN/TP = 2,069/12/4/655.
  • Model: 151,233 parameters, 608,544-byte state artifact.
  • Recorded Apple M3 CPU benchmark: batch-1 p50 0.265417 ms; batch-32 throughput 6582.435206503571 sequences/s, excluding API/network/UI overhead.

This profile deduplicates 575,061 HDFS block traces to 18,383 deterministic (fingerprint, label) representatives before grouped splitting. It is not presented as the conventional all-block benchmark. HDFS supplies binary labels only; NeuralOps does not invent public category or severity targets.

Synthetic product demonstration

The included OpsForge Sim v1 artifact demonstrates anomaly, nine incident categories, and three severity classes. Its locked synthetic test results are binary F1 1.0, category macro-F1 1.0, and severity macro-F1 0.6856184110306733. These reflect authored patterns and are not evidence of real-world generalization.

Engineering included

  • checksum-verified Loghub downloader and leakage-safe grouped/deduplicated splits;
  • statistical baseline and repository-owned packed GRU training loop;
  • masked synthetic category/severity heads with configurable loss weights;
  • validation-selected anomaly threshold and manual-review band;
  • SHA-256-verified predictor, CLI, FastAPI single/batch/sensitivity endpoints;
  • leave-one-event-out score sensitivity with explicit non-causal language;
  • responsive React/TypeScript operator console and reviewed screenshots;
  • non-root, read-only Docker images and health-gated Compose stack;
  • Ruff format/lint, strict mypy, 23 pytest tests, Biome, TypeScript, 4 Vitest tests, 2 Playwright tests, Docker builds and live inference in GitHub Actions;
  • model card, exact experiment JSON, 12-FP/4-FN error analysis, deployment and security documentation.

Artifact attachment

neuralops-opsforge-sim-v1-artifact.tar.gz contains the same small synthetic demo artifact bundled in the repository. Verify the archive using the attached .sha256 file; after extraction, verify every internal file from artifact/SHA256SUMS.

Milestone pull requests

Known limitations

The public benchmark is HDFS-specific and older; probabilities are uncalibrated; arbitrary raw application logs require a domain parser and retraining; synthetic category/severity labels are closed-set; one 128-event model limit applies; no autonomous remediation, hosted deployment, ONNX, Transformer, streaming, or cloud system is included.

Loghub HDFS data is a separate CC BY 4.0 dataset and is not bundled. NeuralOps code and the authored synthetic demo are MIT licensed.

Assets 4
Loading