ToiTenBao/haloprobe-ours-coco-llava15-7b

HaloProbe for LLaVA-1.5-7B

This repository contains the HaloProbe full-method checkpoints and feature caches for COCO and AMBER, arranged as cache/ and checkpoints/ to match the associated TruthPrint release layout.

Contents

• checkpoints/: three HaloProbe full-method MLP checkpoints (seeds 0, 42, and 1337).
• cache/coco/: labeled COCO feature-cache shards used for fitting and holdout evaluation.
• cache/amber/: AMBER feature-cache shards.
• results/results_6.4k.json: per-seed and ensemble results.
• code/: feature extraction and training scripts.

The probe consumes concatenated internal attention/logit features and six external caption-statistics features; the full model input dimension is 2058. See code/haloprobe_cache.py and code/train_haloprobe_full.py for the exact feature construction and model definition.