verdict-typed-base

intfloat/multilingual-e5-base (278M) fine-tuned for 3 epochs on the train split of LocalLLaMA/typed-decisions with the teacher's soft probabilities as targets, the same protocol Laya used for its 0.766. Scores 0.7055 on the 2,000 test decisions (Jev 0.727 zero-shot, Laya typed-decisions 0.766). Published as a reproducible comparison point, not as the default model.

from verdict import Verdict
v = Verdict(model="Manav2op/verdict-typed-base")

Part of Verdict. Training script: train/finetune_typed.py. Full benchmark table: docs/BENCHMARKS.md.

Downloads last month: 11

Safetensors

Model size

0.3B params

Tensor type

F32

Model tree for Manav2op/verdict-typed-base

Base model

intfloat/multilingual-e5-base

Finetuned

(171)

this model

Collection including Manav2op/verdict-typed-base

Verdict: small, fast, honest decision models

Collection

Open Jev/Laya-style decision models with calibrated probabilities and a conformal abstain set. Weights, browser playground, checkpoints. • 4 items • Updated 6 days ago

Evaluation results

accuracy (all) on typed-decisions
test set self-reported

0.706
accuracy (choice) on typed-decisions
test set self-reported

0.682
accuracy (noul) on typed-decisions
test set self-reported

0.803
accuracy (score) on typed-decisions
test set self-reported

0.650