The ablation ladder in Table 5 goes 43.6 → 61.8 → 80.8 ScriptAcc. The first step cleanly isolates Conformer encoder LoRA (+18.2, data held fixed). The last step changes two things at once: LoRA rank 8 → 32 and the addition of the FLEURS-balanced data mix.
The paper states the confound explicitly rather than hiding it, but that costs us the ability to attribute a +19.0 point gain — the single largest jump in the table.
What's needed
Two runs to break the tie:
r=32 encoder LoRA, without the FLEURS mix (original data)r=8 encoder LoRA, with the FLEURS mix
Either one splits the effect. Both give a clean 2x2.
Done when
Table 5 has a row that isolates rank from data, and §7.2 can attribute the gain instead of flagging a confound.
The ablation ladder in Table 5 goes 43.6 → 61.8 → 80.8 ScriptAcc. The first step cleanly isolates Conformer encoder LoRA (+18.2, data held fixed). The last step changes two things at once: LoRA rank 8 → 32 and the addition of the FLEURS-balanced data mix.
The paper states the confound explicitly rather than hiding it, but that costs us the ability to attribute a +19.0 point gain — the single largest jump in the table.
What's needed
Two runs to break the tie:
r=32encoder LoRA, without the FLEURS mix (original data)r=8encoder LoRA, with the FLEURS mixEither one splits the effect. Both give a clean 2x2.
Done when
Table 5 has a row that isolates rank from data, and §7.2 can attribute the gain instead of flagging a confound.