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

@@ -154,14 +154,16 @@ pipeline proceeds as follows:
     previously kept sample $\mathbf{s}'$.
 \end{enumerate}
 
-In addition to deduplication, instruction--response pairs whose
+A length filter is applied to SFT samples based on the IFD
-response contains fewer than 15 tokens are discarded.
+length-bias analysis in Appendix~\ref{sec:ifd_bias}
-Short replies exhibit high per-token variance in both
+(Figure~\ref{fig:length_bias}): instruction--response pairs
-conditional and unconditional loss (Appendix~\ref{sec:ifd_bias},
+whose response contains fewer than 15 tokens are discarded,
-Figure~\ref{fig:length_bias}), which would otherwise distort
+because short replies exhibit both high per-token perplexity
-the IFD-based difficulty estimates used for downstream sample
+($L_{\text{uncond}} \approx 6\text{--}8$, PPL~$\approx 400\text{--}3000$)
-selection.  This length floor is applied per-field, analogous to
+and wide variance in both $L_{\text{cond}}$ and $L_{\text{uncond}}$,
-the pretraining filter described above.
+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}