JokeMaker/mirage-neurips-artifacts

MIRAGE NeurIPS Rebuttal Artifacts

Private research artifacts for the MIRAGE NeurIPS rebuttal experiments.

Contents

activations/
  internvl3_5_4b_hf/
  gemma_3_4b_it/
  qwen3_vl_32b_instruct/
  mimo_vl_7b_rl/
  ovis/<benchmark>/
datasets/
  <model>/contrastive_conversation_pairs.json
  <model>/responses.json
final_results/
  coauthor_final_results.tar.gz
FILES.tsv
MANIFEST.sha256

The four pre-extracted models each contain:

• preextracted_contrastive_features.pt
• preextracted_all_examples_features.pt

Ovis activations are stored as benchmark-specific features_cache.pt files.

Size

Artifact group	Size
Pre-extracted activation stores	81.416 GiB
Ovis activation caches	0.698 GiB
Contrastive datasets	1.652 GiB
All-responses datasets	0.995 GiB
Total payload	approximately 84.8 GiB

Download

Download everything:

hf download JokeMaker/mirage-neurips-artifacts \
  --repo-type dataset \
  --local-dir mirage-neurips-artifacts

Download one model:

hf download JokeMaker/mirage-neurips-artifacts \
  --repo-type dataset \
  --include "activations/qwen3_vl_32b_instruct/*" \
  --include "datasets/qwen3_vl_32b_instruct/*" \
  --local-dir mirage-qwen3-vl

Verify downloaded files:

cd mirage-neurips-artifacts
sha256sum -c MANIFEST.sha256

Access and redistribution

This repository is private and intended for research collaboration. Dataset files may embed or reference benchmark material governed by the original benchmark licenses. Do not make this repository public or redistribute its contents before completing a benchmark-by-benchmark license review.