data_processing

Clustering

This step requires the N×N teacher-ranked matrix to be available at the MODEL_BASE_PATH.

python spectral_clustering4_ibn.py --dataset MSCOCO_i2t --negs hn --nmax 100 --batch_size 32 --K 5

---

Adding Instructions

This step uses the clustered outputs generated from the previous stage.

python modify_posnegtext_instruction2.py

Multiple clustering files with different parameters have already been created and provided in B3/MMEB-train2.
For example,
*_bs32bi_30.130_qwen2b means:

• Teacher model: vlm2vecqwen2b
• Cluster size: $K = 32$
• Parameters: $p = 30$, $m = 100$