Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,101 @@
+---
+license: mit
+task_categories:
+  - visual-question-answering
+  - image-to-text
+tags:
+  - hallucination
+  - vlm
+  - evaluation
+  - diagnostic
+  - representation-analysis
+  - subspace
+pretty_name: "Subspace Validity Suite (SVS)"
+configs:
+  - config_name: default
+    data_files: []
+viewer: false
+---
+
+# Subspace Validity Suite (SVS) — Code & Checkpoints
+
+**Diagnostic toolkit and experimental checkpoints for validating claimed "visual directions" in Vision-Language Models.**
+
+> Associated paper: "The Subspace Validity Suite: Do Visual Directions in Vision-Language Models Survive Basic Sanity Checks?" (WACV 2027, Evaluations and Datasets Track)
+
+## Key Finding
+
+We tested 11 methods for extracting "visual directions" from VLM hidden states across 3 architectures (LLaVA-1.5, LLaVA-Next, BLIP-2), 2 text-only backbones (Vicuna, Mistral), and 2 non-VLM transformer families (BLOOM, deepseek-coder). **Every method fails the gibberish specificity test**: random character sequences activate the claimed visual directions as strongly as visual descriptions (Gib/Vis ≈ 1.00, TOST equivalence p < 0.0001, N = 4,000).
+
+## Repository Structure
+
+### `svs/` — Toolkit Code
+- `svs_toolkit.py` — Complete SVS implementation (6 diagnostic tests)
+- `__init__.py` — Package init
+
+### `experiments/` — Experiment Scripts
+All scripts are self-contained and run on Google Colab with an L4 GPU.
+
+| Script | Experiment |
+|--------|-----------|
+| `scaled_gibberish_cpu.py` | Generate 1000 prompts per type |
+| `scaled_gibberish_gpu.py` | N=4000 gibberish test on Vicuna |
+| `scaled_gibberish_stats.py` | TOST equivalence + Cohen's d |
+| `scaled_makebreak.py` | N=500 makebreak with McNemar's test |
+| `pope_evaluation.py` | POPE evaluation (GRH validation) |
+| `vista_actual.py` | VISTA actual codebase test |
+| `pairwise_discrimination.py` | SVS Test 6: pairwise content classification |
+| `impossible_classification.py` | 36-feature classifier test |
+| `theorem_verification_v2.py` | Eigenvalue dominance verification |
+| `blip2_gibberish.py` | BLIP-2 architecture test |
+| `mistral_gibberish.py` | Mistral backbone test |
+| `generalize_test.py` | Cross-domain (BLOOM + deepseek) |
+
+### `checkpoints/` — Experimental Results
+Raw results for reviewer verification. Load with `json.load()` or `np.load()`.
+
+| Checkpoint | Key Result |
+|-----------|-----------|
+| `gibberish_test/gpu_checkpoint.json` | N=4000 raw alpha values (9 methods × 4 layers × 4 types) |
+| `gibberish_test/statistical_summary.json` | TOST p-values, Cohen's d, bootstrap CIs |
+| `makebreak/checkpoint.json` | N=500 per-image CHAIR results |
+| `pope/pope_checkpoint.json` | POPE accuracy (baseline vs VGCD: +0.000 difference) |
+| `vista_actual/vista_actual_checkpoint.json` | 24-config VISTA gibberish test |
+| `vista_actual/vista_actual_vectors.npz` | Extracted VISTA steering vectors |
+| `classification/pairwise_discrimination.json` | 6-pair accuracy matrix (all >95%) |
+| `theorem/theorem_results.json` | Weight matrix singular value analysis |
+
+### `directions/` — Extracted Subspaces
+Pre-computed PCA/SVD/contrastive directions at layers [8, 16, 24, 32].
+
+| File | Contents |
+|------|----------|
+| `subspaces_9methods.npz` | 9 methods × 4 layers (PCA, image, text, contrastive, etc.) |
+| `directions_vista_nullu.npz` | VISTA + Nullu approximation directions |
+| `subspaces_contrastive.npz` | Contrastive variant directions |
+
+## Quick Verification
+
+```python
+import json, numpy as np
+
+# Verify gibberish test: Gib/Vis ≈ 1.00 for all methods
+with open("checkpoints/gibberish_test/statistical_summary.json") as f:
+    stats = json.load(f)
+for method, result in stats.items():
+    print(f"{method}: Gib/Vis={result['gv_ratio']:.2f}, "
+          f"TOST p={result['tost_p']:.4f}, d={result['cohens_d']:.3f}")
+
+# Verify POPE: VGCD has zero effect
+with open("checkpoints/pope/pope_checkpoint.json") as f:
+    pope = json.load(f)
+bl = pope.get("summary", {}).get("baseline_random", {})
+vg = pope.get("summary", {}).get("vgcd_a1.5_random", {})
+print(f"Baseline POPE: {bl.get('acc', 'N/A')}")
+print(f"VGCD POPE:     {vg.get('acc', 'N/A')}")
+print(f"Difference:    {vg.get('acc', 0) - bl.get('acc', 0):.3f}")
+```
+
+## License
+
+MIT