Update results.py

results.py

@@ -61,7 +61,30 @@ fig_val.update_layout(
 # Show the plot
 validation_loss_graph = fig_val
 
+## lm loss graph
+# Load the data from the cleaned dataset
+data = pd.read_csv('data/lm_loss_txt360_fineweb.csv')  # Replace with your actual file path
 
+# Create the plot
+fig_loss = go.Figure()
+
+# Add TxT360 line
+fig_loss.add_trace(go.Scatter(x=data['Step'], y=data['TxT360'], mode='lines', name='TxT360'))
+
+# Add FineWeb line
+fig_loss.add_trace(go.Scatter(x=data['Step'], y=data['FineWeb'], mode='lines', name='FineWeb'))
+
+# Update layout
+fig_loss.update_layout(
+    title="Loss over Steps: TxT360 vs FineWeb",
+    xaxis_title="Steps",
+    yaxis_title="Loss",
+    legend_title="Models",
+    template="plotly_dark"
+)
+
+# Display the graph
+lm_loss_graph = fig_loss
 
 
 #Perplexity Across Different Buckets (global)
@@ -746,6 +769,7 @@ upsampling_exp = Div(
         H3("Training Evaluations"),
         P("We also conducted full scale training using TxT360 and FineWeb-1.5T. Below are plots of the training and validation loss curves for each dataset. We can see that TxT360 achieves a lower training and validation loss compared to FineWeb-1.5T. "),
         plotly2fasthtml(validation_loss_graph),
+        plotly2fasthtml(lm_loss_graph),
 )
 
 perp1_div = Div(