Böri — Kazakh Missing-Word Reranker (bori-missword)

Stage-2 group reranker for the Kazakh missing-word position task (Böri fill-in-the-blank game).

Metrics

• Reranker validation accuracy (top-1 gap-exact): 0.977
• Honest test-like accuracy: 0.970
• Stage-1 base accuracy: 0.742 → reranker +0.23

Files

• grp_reranker_best.pt — reranker weights + 37 feature names + feat_mean/feat_std (self-contained normalization).
• modeling_reranker.py — GroupReranker (the head) + GapPredictorV13 (Stage-1 base) source.

How to load the reranker

import torch
from modeling_reranker import GroupReranker
obj = torch.load('grp_reranker_best.pt', map_location='cpu', weights_only=False)
feats = obj['features']  # 37 feature names, in order
m = GroupReranker(len(feats), 1024, 256, 512, 4, 320, 8, 0.12)
m.load_state_dict(obj['model'], strict=True); m.eval()

Full pipeline (to score raw sentences)

1. Stage-1 GapPredictorV13 + best_v15.pt → top-10 candidate gaps + 1024-d hidden vectors.
2. word n-gram surprise features (rebuilt from train corpus).
3. masked-LM PLL features (xlm-roberta-large + bert-base-multilingual-cased).
4. normalize 37 features with feat_mean/feat_std, run GroupReranker, argmax over candidates.