Algorithm 1 AST Fine-Tuning with Selective Layer Unfreezing and Early Stopping

Require: Pre-trained AST weights $\Theta$, Fine-tuning dataset ${\mathcal{D}}_{fine}$, Validation dataset ${\mathcal{D}}_{val}$, Evaluation set ${\mathcal{D}}_{eval}$, Patience $P=5$, Max Epochs $N_{max}=30$ Ensure: Optimal Model ${\Theta }^{\ast }$ (Early-stopped), Final Epoch Model ${\Theta }_{final}$ 1:1. Model Initialization & Freezing Mechanism: 2:Set $requires\_grad\leftarrow False$ for all parameters in $\Theta$    ▹ Freeze all layers 3:Unfreeze classification head and final 5 Transformer layers ($B_{8\dots 12}$) 4:Initialize $Los{s}_{best}\leftarrow \infty$, $counter\leftarrow 0$ 5:2. Fine-Tuning and Early Stopping Loop: 6:for each $epoch=1,2,\dots ,N_{max}$ do 7:    $\Theta \leftarrow \mathrm{Train}(\Theta ,{\mathcal{D}}_{fine})$     ▹ Update only 5 layers + classification head 8:    $Los{s}_{val}\leftarrow \mathrm{Validate}(\Theta ,{\mathcal{D}}_{val})$ 9:    if $Los{s}_{val}<Los{s}_{best}$ then          ▹ New global minimum found 10:        $Los{s}_{best}\leftarrow Los{s}_{val}$ 11:        ${\Theta }^{\ast }\leftarrow \Theta$          ▹ Save early-stopped model checkpoint 12:        $counter\leftarrow 0$                ▹ Reset patience tracker 13:    else 14:        $counter\leftarrow counter+1$ 15:        if $counter\ge P$ then             ▹ Patience limit reached 16:           ${\Theta }_{final}\leftarrow \Theta$         ▹ Save state at fine-tuning termination 17:           break                ▹ Early stopping triggered 18:        end if 19:    end if 20:end for 21:3. Model Evaluation: 22:${\mathcal{M}}^{\ast }\leftarrow \mathrm{Evaluate}({\Theta }^{\ast },{\mathcal{D}}_{eval})$        ▹ Evaluate Early-stopped Model 23:${\mathcal{M}}_{final}\leftarrow \mathrm{Evaluate}({\Theta }_{final},{\mathcal{D}}_{eval})$       ▹ Evaluate Final Epoch Model 24:return${\Theta }^{\ast }$, ${\Theta }_{final}$, and Metrics $\{{\mathcal{M}}^{\ast },{\mathcal{M}}_{final}\}$