embeddings-for-preferences-st5-xl

Embeddings for Preferences — ST5-XL

Companion model to "Embeddings for Preferences, Not Semantics".

A sentence embedding whose distances reflect preferential similarity — the probability that a person who agrees with one piece of text agrees with nearby texts — rather than semantic similarity.

This model is a LoRA-adapted sentence-T5-XL (r=16, α=48, q/v projections), fine-tuned with Bradley–Terry loss on counterfactual hard triplets. The LoRA adapter has been merged into the base weights so the model can be loaded as a regular SentenceTransformer.

Why preference instead of semantic similarity?

Two statements with nearly identical wording can take opposite stances on a political or social issue, and a participant who agrees with one will typically disagree with the other. Off-the-shelf embedding models conflate the two — surface similarity dominates the cosine score. This model is fine-tuned to suppress that nuisance signal so cosine similarity tracks who would endorse the same text.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("cartgr/embeddings-for-preferences-st5-xl")

texts = [
    "It should be legal. Religion has no place in politics if we truly want to be free.",
    "It should be illegal. Religion has a place in politics if we truly want to be free.",
    "To ensure true freedom, it's crucial that legislation remains secular, without religious influence.",
]
embs = model.encode(texts, normalize_embeddings=True)

import numpy as np
print("anchor vs distractor (opposite stance, same wording):", embs[0] @ embs[1])
print("anchor vs match     (same stance,    different wording):", embs[0] @ embs[2])

The base ST5-XL would rank the distractor closer than the match (0.87 vs 0.82); this model reverses that (0.79 vs 0.81).

Evaluation

Cosine triplet accuracy across 11 online-deliberation datasets (3 GSC, 3 Remesh, 5 Polis), mean over 5 training seeds:

Per-dataset numbers, baselines (25 embedding models), and the full methodology are in the paper.

Training

• Base encoder: sentence-transformers/sentence-t5-xl
• Adapter: LoRA, r=16, α=48, dropout 0.1, target modules q, v
• Loss: Bradley–Terry over cosine differences log(1 + exp(-(cos(a, p) - cos(a, n))))
• Hyperparameters (val-selected): lr=1.25e-4, n_hard=750, batch size 16, 46 gradient steps
• Training data: 750 synthetic counterfactual hard triplets, generated with GPT-4o from a pool of 2,000 political/social issues sourced from the Habermas Machine corpus and Kialo. The training data was generated independently of the 11 evaluation datasets.

Intended uses and limitations

Intended for policy-style and opinion topics — political and social issues where stances are rooted in values (abortion, immigration, electoral reform, UBI, campus speech, etc.). Generalization to factual/scientific questions, personal-taste preferences, or abstract philosophy is untested. Best on substantive opinion-expressing texts (a sentence or longer); very short comments may not carry enough signal.

Not a stance-detection classifier, not calibrated for absolute agreement probability, and not a substitute for any individual's considered judgment on a binding decision.

Citation

To be added when the corresponding paper is publicly released.

License

Apache-2.0, inherited from the base sentence-T5-XL model.