diff --git a/analysis/hf_reanalysis/D_effective-raw.html b/analysis/hf_reanalysis/D_effective-raw.html
new file mode 100644
index 000000000..389cf20c9
--- /dev/null
+++ b/analysis/hf_reanalysis/D_effective-raw.html
@@ -0,0 +1,14 @@
+
+
+
+
+
+
\ No newline at end of file
diff --git a/analysis/hf_reanalysis/reanalysis.ipynb b/analysis/hf_reanalysis/reanalysis.ipynb
index 712c4bb36..91e3d1dcc 100644
--- a/analysis/hf_reanalysis/reanalysis.ipynb
+++ b/analysis/hf_reanalysis/reanalysis.ipynb
@@ -977,7 +977,7 @@
"id": "e1e91ab4-ae5e-40ff-982e-713925b46922",
"metadata": {},
"source": [
- "Visualize how epoch number affects the Chinchilla Scaling Law"
+ "Visualize how the multi-epoch training pulls away from the Chinchilla scaling law"
]
},
{
@@ -1026,6 +1026,48 @@
" savepath='single-epoch-runs-fitted-multiepoch-D-residuals.html'\n",
")"
]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 42,
+ "id": "1e63c213-4286-468c-a0ea-93d336c98293",
+ "metadata": {
+ "tags": []
+ },
+ "outputs": [],
+ "source": [
+ "def compute_effective_data(params, L, N):\n",
+ " a, b, e, alpha, beta = params\n",
+ " \n",
+ " quot = np.exp(b) / (L - np.exp(e) - np.exp(a)/N**alpha)\n",
+ " \n",
+ " return np.power(quot, 1/beta)"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 43,
+ "id": "831a3693-9b00-4bca-b36e-9df1bf170374",
+ "metadata": {
+ "tags": []
+ },
+ "outputs": [],
+ "source": [
+ "runs = {\n",
+ " **runs, \n",
+ " 'D_effective': compute_effective_data(best_result.x, runs['L'], runs['N'])\n",
+ "}\n",
+ "\n",
+ "scaling_scatter(\n",
+ " runs,\n",
+ " x_key='D',\n",
+ " y_key='R',\n",
+ " z_key='N',\n",
+ " color_key='D_effective',\n",
+ " color_type='log',\n",
+ " savepath='D_effective-raw.html',\n",
+ ")"
+ ]
}
],
"metadata": {