Code for reproducing the experiments in the paper "A Mechanistic Account of Attention Sinks in GPT-2: One Circuit, Broader Implications for Mitigation".

Setup

conda create -n sinks python=3.11 -y
conda activate sinks
pip install -r requirements.txt

Reproducing the Figures and Tables

Figure 1 — Source-Agnostic Shift Histogram (truncated) + Appendix Full Histogram

python experiments_statistical.py --mode bias-term --output-dir results

Outputs:

• results/bias_term_statistical/bq_k_aggregate_plot_truncated.png → Fig 1
• results/bias_term_statistical/bq_k_aggregate_plot.png → Fig 6 (appendix, full histogram)

Figure 2 — EPE-Bias Projection Alignment

python experiments_single_input.py --mode epe-bias-proj --output-dir results

Output:

• results/epe_bias_proj/epe_alignment.png → Fig 2

Figure 3 — EPE Captures the Net Positional Contribution

python experiments_statistical.py --mode epe-validation --output-dir results

Outputs:

• results/epe_validation_statistical/epe_validation_plot.png → Fig 3
• results/epe_validation_statistical/epe_validation_precentiles (numerical values for experiments)

Figure 4 — Coordinate-Level Alignment Histogram (truncated) + Appendix Full Histogram

python experiments_statistical.py --mode coord-alignment --output-dir results

Outputs:

• results/coord_alignment_statistical/coord_alignment_histogram_truncated.png → Fig 4
• results/coord_alignment_statistical/coord_alignment_histogram.png → Fig 8 (appendix, full histogram)

Figure 5 — Intervention Attention Maps

python intervention_analysis.py --mode sentence --output-dir results

Outputs:

• results/sentence_analysis/layer_04_avg.png through layer_11_avg.png → Fig 5 (layers 4--11)

Table 1 — BOS Attention Statistics

python intervention_analysis.py --mode dataset --output-dir results

Outputs:

• results/dataset_analysis/bos_attention_summary_mid_layers.txt → Table 1
• results/dataset_analysis/bos_attention_summary_mid_layers.csv

Figure 7 (appendix) — Massive Activations in EPE_1

python experiments_single_input.py --mode massive-activations --output-dir results

Output:

• results/massive_activations/massive_activations_in_ppe.png → Fig 7