belief-state-probes-patching-v2

CANONICAL N1 patching run with 6 conditions incl. the realized-size-matched random control (seed 42, k=2-5, CPU; reproduces the v1 GPU quantities to 2.3e-7). HEADLINE: belief clamp statistically indistinguishable from realized-size-matched random perturbations (belief-minus-control CI includes zero at k>=3, -0.007 [-0.014,-0.000] at k=2), never better at any horizon. Per-pair achieved ratios + convergence flags recorded (1084/1146 within 2%).

Dataset Info

• Rows: 1148
• Columns: 2

Columns

Column	Type	Description
source_file	Value('string')	originating JSON file within the results dir
record	Value('string')	full JSON record (summary dict or one list element), no truncation

Generation Parameters

{
  "script_name": "scripts/upload_artifacts.py",
  "model": "toy transformer (Shai et al. 2024 arch: 4L x 1H, d_head 8, d_model 64)",
  "description": "CANONICAL N1 patching run with 6 conditions incl. the realized-size-matched random control (seed 42, k=2-5, CPU; reproduces the v1 GPU quantities to 2.3e-7). HEADLINE: belief clamp statistically indistinguishable from realized-size-matched random perturbations (belief-minus-control CI includes zero at k>=3, -0.007 [-0.014,-0.000] at k=2), never better at any horizon. Per-pair achieved ratios + convergence flags recorded (1084/1146 within 2%).",
  "experiment_name": "belief-state-probes",
  "job_id": "local:cpu",
  "cluster": "local",
  "artifact_status": "final",
  "canary": false,
  "hyperparameters": {
    "see": "record column; full configs in metrics.json rows"
  },
  "input_datasets": []
}

Usage

from datasets import load_dataset

dataset = load_dataset("latkes/belief-state-probes-patching-v2", split="train")
print(f"Loaded {len(dataset)} rows")

---