Archive: earlier models (reference only)
Do not use these files with hybrid_engine/. The current model is v2.0 in hybrid_engine/artifacts/ (see the main README). The models here were trained on different inputs (they use community name and flair, user/post IDs or feed position) and have their own vocabularies and standardisation. They are kept so the results in REPORT.md can be checked and compared.
| Folder | What it is | Status |
|---|---|---|
engine_v1_1/ |
the previous hybrid engine (same design, but its models used community name + flair). ONNX, meta.json, reports, pytorch_checkpoints/ (5 seeds per model) |
replaced by v2.0 |
phase1_models/ |
the first five models (LR, DeepFM, Wide&Deep, DCN V2, SASRec), trained with user/post IDs and feed position. ONNX, checkpoints, metadata, Model_Explainer.docx, example_inference.py |
research baseline |
v1_two_stage_engine/ |
Stage-2 models of the first two-stage engine (tag match, then model). ONNX, checkpoints, Stage-1 weights and results | research baseline, outperformed |
The engine v1.1 files can be loaded on their own with HybridRecommender(artifacts_dir="archive/engine_v1_1"), but that version needs content vectors built with its own text recipe ("title. community. flair. body"), not the v2.0 one.