LoneWolfgang/telelogs-rca-classifier

TeleLogs RCA — Numeric + Text Fusion (confidence_concat)

8-way root-cause-analysis classifier (C1–C8) for netop/TeleLogs 5G drive-test traces. A numeric MLP branch and a MiniLM text branch are fused with a confidence_concat head. The model owns its full front end: give it the raw drive-test question string and it parses, featurizes, verbalizes, and predicts.

Usage

from transformers import AutoModel
model = AutoModel.from_pretrained("LoneWolfgang/telelogs-rca-classifier", trust_remote_code=True)

label = model.predict(raw_question_string)          # -> "C3"
labels = model.predict([q1, q2, q3])                # -> ["C1", "C7", ...]
probs = model.predict_proba(raw_question_string)    # -> np.ndarray (8,)

Or via a pipeline:

from transformers import pipeline
pipe = pipeline("telelogs-rca", model="LoneWolfgang/telelogs-rca-classifier", trust_remote_code=True)
pipe(raw_question_string)        # -> {"label": "C3", "score": 0.98}

Test-set performance

confidence_concat fusion, held-out test split. See metrics.json.

Notes

All preprocessing (numeric imputer + scaler, the two branch-feature scalers, the z-score verbalizer, the label encoder) is fit on train only and bundled with the weights, so predictions reproduce the offline evaluation. A trace that fails to parse returns a uniform-probability fallback rather than raising.