Atlas Recommendation Engine — Model & Feature Artifacts

This repository is a private model repository hosting the pre-trained recommendation model artifacts and behavioral feature tables for the Atlas ML Inference Microservice.

Repository Overview

Service: Atlas External ML Inference Microservice
Active Model Version: production_v1
Visibility: Private (Requires authenticated HF_TOKEN to retrieve)
Runtime Consumer: Railway ML Inference Service container

Artifact Categories

1. Online Production Inference (Required)

These artifacts power the active two-stage online recommendation and re-ranking pipeline:

models/production_v1/item_similarity.pkl (~10.8 MB): Sparse TF-IDF item-item co-visitation matrix for recall candidate generation (132K items, 317K pairs).
models/production_v1/lightgbm_ranker.txt (~358 KB): LambdaRank LightGBM model trained on 16 behavioral features.
models/production_v1/popularity_baseline.pkl (~5.6 MB): Pre-computed popularity baseline for fallback.
features/retailrocket/item_features.parquet (~5.9 MB): Pre-computed item interaction and conversion feature store.
features/retailrocket/user_features.parquet (~30.8 MB): Pre-computed user engagement and session feature store.

2. Swagger / Testing & Offline Evaluation

models/production_v1/svd_model.pkl (~173.2 MB): Matrix factorization model preserved for interactive Swagger testing, offline evaluation, and experimentation. Not required for the primary production online path.

3. Metadata & Manifests

models/promoted_model.json: Active model pointer (production_v1).
models/production_v1/artifact_manifest.json: SHA-256 cryptographic checksums and schema requirements.
models/production_v1/candidate_models_metadata.json: Model component hyperparameters and dimensions.
models/production_v1/ranker_metadata.json: Feature importance and ranking configuration.

Directory Structure

.
├── README.md
├── models/
│   ├── promoted_model.json
│   └── production_v1/
│       ├── artifact_manifest.json
│       ├── candidate_models_metadata.json
│       ├── item_similarity.pkl
│       ├── lightgbm_ranker.txt
│       ├── popularity_baseline.pkl
│       ├── ranker_metadata.json
│       └── svd_model.pkl
└── features/
    └── retailrocket/
        ├── item_features.parquet
        └── user_features.parquet

Runtime Consumption

The Railway ML service container downloads these artifacts during container startup using huggingface_hub when provided with HF_MODEL_REPO and HF_TOKEN:

snapshot_download(
    repo_id=HF_MODEL_REPO,
    local_dir=ARTIFACTS_PATH,
    token=HF_TOKEN
)

SHA-256 integrity and schema compatibility are validated against artifact_manifest.json before serving starts.

Downloads last month: -; Downloads are not tracked for this model. How to track