doc: add SFT length-filter rationale — high PPL + high variance from A.3 IFD figure

- Add 15-token length floor to SFT samples, with explicit reference
  to Appendix A.3 / Figure 5 (ifd_length_grid)
- Short replies (<10 tokens) show both high per-token perplexity
  (L_uncond ~6-8, PPL ~400-3000 vs long replies ~2-3, PPL ~7-20)
  and wide variance in L_cond (span 0-17.5), which distorts
  downstream IFD-based difficulty estimates.

main.tex

@@ -154,8 +154,16 @@ pipeline proceeds as follows:
     previously kept sample $\mathbf{s}'$.
 \end{enumerate}
 
-An optional LLM-as-Judge scoring module provides multi-dimensional
+A length filter is applied to SFT samples based on the IFD
-quality scores that can be used to filter low-quality samples.
+length-bias analysis in Appendix~\ref{sec:ifd_bias}
+(Figure~\ref{fig:length_bias}): instruction--response pairs
+whose response contains fewer than 15 tokens are discarded,
+because short replies exhibit both high per-token perplexity
+($L_{\text{uncond}} \approx 6\text{--}8$, PPL~$\approx 400\text{--}3000$)
+and wide variance in both $L_{\text{cond}}$ and $L_{\text{uncond}}$,
+which would distort downstream IFD-based difficulty estimates.
+The threshold is applied per-field, analogous to the pretraining
+filter described above.
 
 \subsection{DPO Data Generation}
 
@@ -302,7 +310,7 @@ Sequence length & 2,048 tokens \\
 \centering
 \includegraphics[width=0.50\linewidth]{data/loss_compare.png}
 \caption{Training loss curves: GPT-2 residual scaling vs.~Kaiming
-initialization over $\sim$20B tokens.}
+initialization over $\sim$5B tokens.}
 \label{fig:loss}
 \end{figure}
 
@@ -547,17 +555,6 @@ spread.}
 \label{fig:ckpt_weight_density}
 \end{figure}
 
-\begin{figure}[H]
-\centering
-\includegraphics[width=0.95\linewidth]{data/ckpt_weight_density_per_run.png}
-\caption{Per-checkpoint weight density breakdowns. The three checkpoints
-are \texttt{norm-15bt} (AdamW, Normal init, 15B tokens), \texttt{kami-15bt}
-(AdamW, GPT-2 residual scaling, 15B tokens), and \texttt{muon-25bt}
-(Muon, 25B tokens). Note: the iteration labels inside the figure reflect
-legacy script metadata and should be ignored in favour of the token
-budgets stated here.}
-\label{fig:ckpt_weight_density_per_run}
-\end{figure}
 
 % ======================================================================
 \section{Conclusion}