thiagochris/cbcc

v29NA-openvino-int8-serving

Multi-dimensional classifier for central-bank communications produced by the CBCommunication training pipeline (multi_task_gist rung).

Provenance

Field	Value
Trainer	multi_task_gist
Model kind	multi_task
Encoder body	Qwen/Qwen3-Embedding-4B
Loss	cached_gist
Taxonomy version	2026-04-rev2 (sha256 e7c237aac8db66ca)
Training examples	3584
Validation examples	1809
Git commit	9d90b862 (dirty)
Created	2026-04-28T02:41:30.780109+00:00

Dimensions and labels

topic (21 classes)

• Climate change
• Crisis management
• Currency circulation and management
• Financial inclusion
• Financial stability
• Fiscal policy
• Governance
• MP - balance sheet size and asset purchase programs
• MP - credit
• MP - economic activity
• MP - exchange rate
• MP - inflation
• MP - interest rate
• MP - labor market
• MP - open market operations
• MP - reserve requirements
• Metadata
• Payment system
• Structural economic reform
• Supervision and regulation
• Technological innovation and fintech

temporal_orientation (2 classes)

• Backward-looking
• Forward-looking

audience (6 classes)

• Business Sector
• Financial Sector
• General Public
• Government
• International Stakeholders
• Metadata

sentiment (6 classes)

• Confidence-building
• Dovish
• Hawkish
• Neutral/Balanced
• Not applicable
• Risk-highlighting

Evaluation (held-out validation set)

Dimension	Macro F1
topic	0.8008
temporal_orientation	0.8957
audience	0.7590
sentiment	0.7511

Intended use

Classify sentences from central-bank speeches, press releases, and financial-stability reports along the four CBC taxonomy dimensions (topic, temporal orientation, audience, sentiment). Produced for research and policy analysis at the IMF.

Limitations

• Trained on a small labeled set; tail classes (low support) carry less reliable per-class metrics.
• Multilingual coverage depends on the encoder and labeled-set coverage; the current Qwen3-Embedding family is strong cross-lingually, but performance still varies on low-resource languages and OCR-heavy inputs.
• Sentiment / temporal labels reflect the taxonomy decision rules in the source workbook; downstream consumers should re-read those rules before interpreting per-class deltas.

How to load

# Recommended after registering this revision in config/classifiers.toml:
from cb_communication.processing.classification import load_named_classifier

# Auto-resolves the artefact via the registry's pinned hub_revision.
# Private collaborators authenticate with their own HF_TOKEN.
with load_named_classifier("v29NA-openvino-int8-serving") as clf:
    results = clf.classify_chunk([
        "Inflation expectations remain anchored at 2 percent.",
    ])

Alternative — explicit tag-pinned Hub load:

from huggingface_hub import snapshot_download
from cb_communication.processing.classification import load_classifier

local = snapshot_download(repo_id="thiagochris/cbcc", revision="v29NA-openvino-int8-serving")
with load_classifier(local) as clf:
    ...

Both call paths satisfy the canonical MultiTaskClassifier Protocol (see cb_communication/processing/classification/multi_task_classifier.py) — the runtime dispatches on model_kind from manifest.json.