Skip to content

A straightforward method for training your LLM, from downloading data to generating text.

Correct equation format in objectives.md

Mr. Ohucebcc2dee12c
diff --git a/docs/foundations/objectives.md b/docs/foundations/objectives.md
index 8dae25c..5658887 100644
--- a/docs/foundations/objectives.md
+++ b/docs/foundations/objectives.md
@@ -94,8 +94,7 @@ So an untrained model often starts near that value.
SFT still uses next-token cross-entropy, but only assistant completion tokens count:
\[
-\mathcal{L}_{\text{SFT}}
-=
+\mathcal{L}_{\text{SFT}} =
\frac{\sum_{b,t} m_{b,t}\,\ell_{b,t}}
{\sum_{b,t} m_{b,t}}
\]
@@ -155,8 +154,7 @@ against a frozen reference model:
\]
\[
-\mathcal{L}_{\text{DPO}}
-=
+\mathcal{L}_{\text{DPO}} =
-\log \sigma\left(\beta(\Delta_\pi - \Delta_{\text{ref}})\right)
\]
@@ -187,8 +185,7 @@ r_t(\theta) =
The clipped policy objective is:
\[
-\mathcal{L}_{\text{PPO}}
-=
+\mathcal{L}_{\text{PPO}} =
-\mathbb{E}_t
\left[
\min