calerio/silent-signals-rqb

Multiclass ingroup classifier (17 classes; train-only label space). RoBERTa-base.

Headline metric: f1_macro — macro-F1 on the grouped-split test (held-out roots; see docs/rq_b_report.md § 4.1).

Variants

Each variant is checked into its own branch. Load with:

from transformers import AutoModelForSequenceClassification, AutoTokenizer
model = AutoModelForSequenceClassification.from_pretrained("calerio/silent-signals-rqb", revision="<branch>")
tok   = AutoTokenizer.from_pretrained("calerio/silent-signals-rqb", revision="<branch>")

Branch	Variant id	f1_macro	Notes
term-seed123	rqb_term_seed123	0.3626	default
term-seed42	rqb_term_seed42	0.3491	—
term-seed7	rqb_term_seed7	0.3468	—
term-altsplit-seed42	rqb_term_altsplit_seed42	0.9964	raw leader (not default — see below)
term-weighted-seed42	rqb_term_weighted_seed42	0.3346	—
text-only-seed42	rqb_text_only_seed42	0.3140	—

Default variant

rqb_term_seed123 — see the per-task default_variant_rationale in data/manifests/model_inventory.json of the project repo.

Restricted to Run A (term arm, grouped split, no run_tag) — the headline system in docs/rq_b_report.md. Run C (altsplit, 0.996) is excluded: it's a glossary-determinism artefact, not generalization, and defaulting to it would misrepresent RQ-B's central finding. The Run C model remains selectable in the dropdown for the educational tab (Step 6).

Where this came from

Bocconi 597 NLP group project on dog-whistle detection and disambiguation, on the silent_signals corpus (Kruk et al. 2024). Full methodology: docs/DESIGN_DEFENSE.md + per-RQ reports in the project repo. HF Space build write-up: docs/hf_space_report.md.