Mantis-8M TelecomTS RCA head
The newly trained RCA head based on
notebooks/telecomts_model_evaluation.ipynb classifies anomalous 128-step,
18-channel TelecomTS KPI windows. It uses the paris-noah/Mantis-8M backbone,
which the serving image bakes in offline, plus the task checkpoint
mantis_pretrained_rca.pt. Recovery used seed 42 and 35 epochs. No retraining
is part of serving.
Input channel order and the ordered RCA class mapping are recorded in
rca_preprocessing.json. Protocols map
TCP→0, UDP→1, and missing/None/empty→2. For RCA, the TelecomTS reference computes
each channel's mean and population standard deviation across all RCA examples
and timesteps in a split. The service uses the fixed training split values
from rca_preprocessing.json; it never normalizes each request window. The
original notebook independently normalizes its test split, so the file also
records test statistics for exact notebook parity. The pretrained encoder
linearly interpolates 128 steps to 512 with align_corners=False.
Class indices, in order: 0 Antenna Failure; 1 Co-Channel Interference (Mild);
2 Co-Channel Interference (Severe); 3 Faulty RF Filters (Temporal);
4 Doppler Shift (Severe); 5 Faulty Handover Algorithm (Too Frequent);
6 Buffer Overflow (Gradual Buildup); 7 Resource Allocation Bugs;
8 High Network Congestion (Gradual Buildup); 9 High Network Congestion
(Sudden Spike). These match the TelecomTS anomaly_type_2_id mapping.
See the TelecomTS reference preprocessing and class mapping.
The notebook SHA-256 is
7710e3fbd47e8a7237ac1cef1629b8c844d08b855012d837daef64de320c3cc4.
The dataset revision is 1b3a88a440f4922edfcd1e230e42c003c0e13bdf
with split seed 42. The backbone revision used by AD parity is
bc7d5ab40c02133386a28e2c127f35c17c86901d. The preprocessing file
SHA-256 is 1a8e0144bb4a65f48133c4ad36491aba6432b817d64b8d3521a3c8de913ddeae.
Recovery result (issue #196)
The new 35-epoch checkpoint was evaluated through the existing predictor on the original 80/20 seed-42 split. On all 192 held-out 10-class RCA examples, training-split normalization gave 92.71% accuracy and 0.9274 macro F1, passing both 0.90 gates. The notebook-style test-split normalization comparison gave 92.19% accuracy and 0.9086 macro F1. These are scores for the new checkpoint, not the unavailable historical weights. The 58 Jamming examples were reported separately as unsupported inputs, with no accuracy claim. Full per-class precision, recall, F1, support, confusion matrix, predictions, and provenance are in the validation report.
Checkpoint SHA-256: 0ecd02d50eed50130476c0c543503f9f3d1595507502664d3e5b8fb01c3eccfe.
Preprocessing SHA-256: 1a8e0144bb4a65f48133c4ad36491aba6432b817d64b8d3521a3c8de913ddeae.
Both files are retained privately at
~/Library/Application Support/ai-driven-network-remediation/issue-196/,
independent of the training directory and repository workspace. A workspace
copy at model-serving/ran-ml-service/.private-artifacts/issue-196/ is ignored
by Git and excluded from image builds. Both copies have matching hashes; the
persistent copy was reloaded through the classifier.
Synthetic TelecomTS results alone do not establish production generalization.
Pinned serving reference coverage
tests/fixtures/rca_reference.json records local /v1/classify outputs and
ten logits for four TelecomTS fixtures using the recovered checkpoint. It also
pins each fixture's SHA-256, the notebook hash, the dataset revision and split
seed, the ordered ten-class mapping, and both artifact hashes. The dedicated
model-serving parity CI compares class index and label exactly when
RCA_ARTIFACT_COMMIT is set to the published repository commit, with absolute
tolerances of 0.0001 for confidence and logits. These are local serving
observations, not outputs from the unavailable historical notebook checkpoint.
| Fixture | Class index | Confidence |
|---|---|---|
| Antenna failure | 0 | 0.9293 |
| Severe co-channel interference | 2 | 0.9502 |
| Severe Doppler shift | 4 | 0.9568 |
| Sudden high congestion | 9 | 0.8996 |