Attuned Resonance Intake — Multi-Head RoBERTa for Call-Center Intake
Six-head classifier/regressor built on roberta-base. Given a call transcript, emits a structured intake record used downstream by the outcome predictor and the PPO router.
The GitHub repo slug is still "CEPM" during the gradual rename; the Hugging Face slugs were migrated to attuned-resonance-* on 2026-05-09 (HF preserves the old cepm-* URLs as redirects).
Research/educational use only. See disclaimer below.
Heads
Head
Type
Output
intent
Softmax (8 classes)
Caller domain (zip-source vocabulary)
action
Softmax (7 classes)
Call type in the predictor's call_type vocabulary — bridges the cascade NLP→predictor
The action head exists to fix a silent cascade bug: the intent head emits zip-source labels (above), but the downstream predictor's _INTENT_MAP is keyed on the synthetic call_types it was trained against. Without the action head, every NLP intent was silently mapped to predictor class 0. The action head emits in the predictor's native vocabulary so the cascade carries a feature the predictor actually understands.
Indices match generator/config.py:call_types (do not re-sort):
billing
technical
general
account
complaint
cancellation
upgrade
Silver-label heuristic + class distribution
Action labels are silver-standard, generated by a deterministic keyword heuristic (_action_label_for in models/intake/dataset.py) with priority order: cancellation > upgrade > technical > billing > account > complaint > general. Distribution on the 90,093-record corpus:
Class
Share
Notes
general
56%
Corpus-driven (Medicare ≈ 68% of records, mostly coverage clarifications)
billing
20%
cancellation
11%
technical
7%
upgrade
5%
"premium" keyword removed after 82% false-positive rate on insurance_outbound (matched "your premium will be …")
complaint
1%
account
1%
Imbalance is corrected at training time by a sqrt-tempered inverse-frequency class-weight vector applied to action_CE:
Linear 1/n would give account a ~100× weight vs general — too aggressive for silver labels.
Sqrt-tempered (mean-normalized) compresses that to ~6.6× (general 0.29 vs account 1.97), correcting the imbalance without overfitting to whichever heuristic mistakes happen to land in the rare classes.
Tokenizer: RobertaTokenizerFast (Rust-backed, byte-identical to slow)
Tracked: MLflow experiment prod-intake-b2
Metrics (best checkpoint, epoch 4)
Metric
Value
train_loss (sum of 7 head losses)
5.5984
val_loss
5.6088
Best val_loss
5.6088
Wall clock
~25 min total (~22 min training, ~3 min upload)
Train ≈ val throughout (gap < 0.01 at epoch 4) → no overfit signal at 4 epochs. Per-epoch deltas were shrinking (-0.07, -0.04, -0.03, -0.03), indicating the model was approaching convergence; further epochs at decayed LR would likely buy diminishing returns.
The combined loss adds the action head's weighted CE on top of the prior 5-head sum, so the absolute loss is ~1.5–2 units higher than the pre-B2 5-head run (3.99 val); the shape (monotone decrease, train≈val) is the like-for-like comparison.
Intended Use
Research on multi-head transformer fine-tuning + cascade-aware vocabulary design
Educational demonstrations of NLP → downstream model cascades
Benchmarking against the synthetic-data environment in this repo
Out of Scope / Not Intended
Any production or commercial use. Not validated for operational deployment.
Customer-facing decision-making. The archetype heads (Jung/Campbell) are exploratory and have low calibrated confidence.
Any setting where misclassification carries material risk.
Limitations
Silver-labeled action head. The action labels come from a keyword heuristic, not human annotation. Class-weight tempering helps, but the head's accuracy is bounded by heuristic accuracy on the rare classes.
Archetype heads remain speculative. Confidence is typically <15% on real samples. Treat as exploratory.
English-only. Transcripts were US English with finance/medicare/automotive bias; no telecom data in the current corpus.
Convergence headroom. 4 epochs at LR 2e-5 was a deliberate stop after the elbow of the loss curve; longer runs at decayed LR may improve val loss slightly.
Future Work
Telecom call-type vocabulary. When the corpus expands to telecom transcripts, the action head's vocabulary will migrate to a customer-facing taxonomy: Internet, TV, Mobile, Cancel, Retention, Billing, Other. The existing 7-class call_type vocab (billing/technical/general/account/complaint/cancellation/upgrade) is operationally accurate for the current insurance/medicare/automotive corpus but would feel awkward for telecom. Tracked in project_v21_plan.md.
Predictor refresh against the new intake feature distribution (action-channel cascade is now active; predictor was trained against the old intent-channel fallback).
Router (PPO) training to close the v2.1 three-model loop.
Changelog
2026-04 — B2 retrain (this card)
6th head added (action, 7-class) bridging the cascade NLP→predictor. Architecture + silver-label heuristic in commit 4933d68; trainer + class weights in 5f517b4; pooler removal for DDP in 1605261.
Retrained on 2× H100 SXM via accelerate launch --num_processes=2. 4 epochs, max_length=512, effective batch 64. Best val_loss 5.6088.
Sqrt-tempered inverse-frequency class weights applied to action loss to correct the 56%/1% imbalance without overfitting silver-label noise.
Cascade switch in notebooks/research.ipynb Section 9b now feeds the predictor intake["action"] (call_type vocab) instead of intake["intent"] (zip-domain vocab); the latter silently fell back to class 0 inside the predictor.
2026-04 — clean-schema retrain
Replaced the per-domain .get(domain, 0) label fallback with a strict whitelist (_ZIP_TO_INTENT) and INTENT_LABELS = sorted(set(...)). Previously, three malformed/duplicate zips were silently mislabeled as class 0 (label bleed).
Excluded zips: a typo-and-mixed-domain bundle (home_ervice_inbound&telecom _outbound) and two (reupload) duplicates whose transcript counts matched the originals.
