diff options
Diffstat (limited to 'notebooks')
| -rw-r--r-- | notebooks/build_notebook.py | 9 | ||||
| -rw-r--r-- | notebooks/recursive_reasoning_chaos.ipynb | 33 |
2 files changed, 24 insertions, 18 deletions
diff --git a/notebooks/build_notebook.py b/notebooks/build_notebook.py index 8a525b2..930dc68 100644 --- a/notebooks/build_notebook.py +++ b/notebooks/build_notebook.py @@ -35,7 +35,10 @@ md("## 0. Setup") code("""%pip install -q torch einops pydantic huggingface_hub numpy matplotlib tqdm import numpy as np, matplotlib.pyplot as plt, torch from tqdm.auto import tqdm -print("torch", torch.__version__, "| cuda", torch.cuda.is_available())""") +print("torch", torch.__version__, "| cuda", torch.cuda.is_available()) +if not torch.cuda.is_available(): + print("\\n⚠️ No GPU detected — the JVP/rollout cells will be SLOW on CPU.") + print(" Colab: Runtime → Change runtime type → Hardware accelerator → GPU (T4), then re-run.")""") md(f"""## 1. Load a trained model from HuggingFace @@ -126,7 +129,7 @@ def leading_ftle(inp, lab, pid, n=128, n_seg=16, seed=0): ok=(((inner.lm_head(zH)[:,pe:].float().argmax(-1)==Y)|~m).all(-1)).cpu().numpy() return ftle, ok -ftle, succ = leading_ftle(inp, lab, pid, n=128) +ftle, succ = leading_ftle(inp, lab, pid, n=64) # n=64 already separates cleanly; raise for tighter histograms print(f"success rate {succ.mean():.2f} | median λ1 success {np.median(ftle[succ]):+.4f} vs failure {np.median(ftle[~succ]):+.4f}") print(f"AUC(-λ1 -> success) = {auc(-ftle, succ.astype(int)):.3f} (>0.5 means failures are more chaotic)") plt.figure(figsize=(6,4)) @@ -163,7 +166,7 @@ code("""def extended_rollout(inp, lab, pid, n=256, n_seg=128, seed=0): zH=Hmod(zH, zL, **si) p=inner.lm_head(zH)[:,pe:].float().argmax(-1) EX.append(((p==Y)|~m).all(-1).float().cpu().numpy()) - DR.append((torch.zeros(n) if prev is None else (zH-prev).float().flatten(1).norm(1).cpu()).numpy()) + DR.append((torch.zeros(n) if prev is None else (zH-prev).float().flatten(1).norm(dim=1).cpu()).numpy()) prev=zH.detach() return np.stack(EX,1), np.stack(DR,1) diff --git a/notebooks/recursive_reasoning_chaos.ipynb b/notebooks/recursive_reasoning_chaos.ipynb index 1f00628..c9e96a9 100644 --- a/notebooks/recursive_reasoning_chaos.ipynb +++ b/notebooks/recursive_reasoning_chaos.ipynb @@ -2,7 +2,7 @@ "cells": [ { "cell_type": "markdown", - "id": "f991b29d", + "id": "9ed15c04", "metadata": {}, "source": [ "# Recursive Reasoning Failures are Chaotic\n", @@ -27,7 +27,7 @@ }, { "cell_type": "markdown", - "id": "bb82e0a8", + "id": "f9d8321d", "metadata": {}, "source": [ "## 0. Setup" @@ -36,19 +36,22 @@ { "cell_type": "code", "execution_count": null, - "id": "89a96776", + "id": "ea63e86f", "metadata": {}, "outputs": [], "source": [ "%pip install -q torch einops pydantic huggingface_hub numpy matplotlib tqdm\n", "import numpy as np, matplotlib.pyplot as plt, torch\n", "from tqdm.auto import tqdm\n", - "print(\"torch\", torch.__version__, \"| cuda\", torch.cuda.is_available())" + "print(\"torch\", torch.__version__, \"| cuda\", torch.cuda.is_available())\n", + "if not torch.cuda.is_available():\n", + " print(\"\\n⚠️ No GPU detected — the JVP/rollout cells will be SLOW on CPU.\")\n", + " print(\" Colab: Runtime → Change runtime type → Hardware accelerator → GPU (T4), then re-run.\")" ] }, { "cell_type": "markdown", - "id": "bbd25841", + "id": "1da5260b", "metadata": {}, "source": [ "## 1. Load a trained model from HuggingFace\n", @@ -61,7 +64,7 @@ { "cell_type": "code", "execution_count": null, - "id": "f6a83ba0", + "id": "d1821039", "metadata": {}, "outputs": [], "source": [ @@ -94,7 +97,7 @@ }, { "cell_type": "markdown", - "id": "086e411f", + "id": "e0b10696", "metadata": {}, "source": [ "## 2. The core result — failures are more chaotic (leading FTLE / λ₁)\n", @@ -112,7 +115,7 @@ { "cell_type": "code", "execution_count": null, - "id": "791d02dc", + "id": "01a845b8", "metadata": {}, "outputs": [], "source": [ @@ -164,7 +167,7 @@ " ok=(((inner.lm_head(zH)[:,pe:].float().argmax(-1)==Y)|~m).all(-1)).cpu().numpy()\n", " return ftle, ok\n", "\n", - "ftle, succ = leading_ftle(inp, lab, pid, n=128)\n", + "ftle, succ = leading_ftle(inp, lab, pid, n=64) # n=64 already separates cleanly; raise for tighter histograms\n", "print(f\"success rate {succ.mean():.2f} | median λ1 success {np.median(ftle[succ]):+.4f} vs failure {np.median(ftle[~succ]):+.4f}\")\n", "print(f\"AUC(-λ1 -> success) = {auc(-ftle, succ.astype(int)):.3f} (>0.5 means failures are more chaotic)\")\n", "plt.figure(figsize=(6,4))\n", @@ -178,7 +181,7 @@ }, { "cell_type": "markdown", - "id": "25aa2620", + "id": "36e16b8f", "metadata": {}, "source": [ "## 3. Why — transient chaos: failures *escape* with more compute\n", @@ -192,7 +195,7 @@ { "cell_type": "code", "execution_count": null, - "id": "106f13d3", + "id": "6603b85d", "metadata": {}, "outputs": [], "source": [ @@ -216,7 +219,7 @@ " zH=Hmod(zH, zL, **si)\n", " p=inner.lm_head(zH)[:,pe:].float().argmax(-1)\n", " EX.append(((p==Y)|~m).all(-1).float().cpu().numpy())\n", - " DR.append((torch.zeros(n) if prev is None else (zH-prev).float().flatten(1).norm(1).cpu()).numpy())\n", + " DR.append((torch.zeros(n) if prev is None else (zH-prev).float().flatten(1).norm(dim=1).cpu()).numpy())\n", " prev=zH.detach()\n", " return np.stack(EX,1), np.stack(DR,1)\n", "\n", @@ -234,7 +237,7 @@ }, { "cell_type": "markdown", - "id": "306d11c9", + "id": "32c11219", "metadata": {}, "source": [ "## 4. Basin accessibility — input-determined or initial-condition-determined?\n", @@ -249,7 +252,7 @@ { "cell_type": "code", "execution_count": null, - "id": "682927dd", + "id": "53a705d1", "metadata": {}, "outputs": [], "source": [ @@ -296,7 +299,7 @@ }, { "cell_type": "markdown", - "id": "22ef7f6f", + "id": "6ca89f3b", "metadata": {}, "source": [ "## What this shows\n", |