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.mdindex 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