Finetune BGE-M3

[tenafrangelos]

How could I finetune dense and sparse embedding only ?
I try to use this script :

%%bash
torchrun --nproc_per_node 1 \
	-m FlagEmbedding.finetune.embedder.encoder_only.m3 \
	--model_name_or_path /home/alex/ejada/developers/martina/my_cache/models--BAAI--bge-m3 \
    --cache_dir ./cache/model \
    --train_data ./ft_data/training.json \
    --train_group_size 4 \
    --query_max_len 256 \
    --passage_max_len 256 \
    --pad_to_multiple_of 4 \
    --query_instruction_for_retrieval 'Represent this sentence for searching relevant passages: ' \
    --query_instruction_format '{}{}' \
    --knowledge_distillation False \
	--output_dir ./test_encoder \
    --learning_rate 1e-5 \
    --num_train_epochs 1 \
    --per_device_train_batch_size 1 \
    --dataloader_drop_last True \
    --warmup_ratio 0.1 \
    --logging_steps 1 \
    --save_steps 1000 \
    --negatives_cross_device \
    --temperature 0.02 \
    --sentence_pooling_method cls \
    --normalize_embeddings True \
    --kd_loss_type m3_kd_loss \
    --unified_finetuning True \
    --use_self_distill True \
    --fix_encoder True \
    --colbert_dim 0 \
    --self_distill_start_step 0

[545999961]

ColBERT vector and sparse embedding are finetuned together. If you want to remove the colbert vector, you need to remove the code in the finetune module.

[tenafrangelos]

Thanks for your answer . It's work.
I update loss function to the following :
Before :

1. return dense_scores + 0.3 * sparse_scores + colbert_scores
2. loss = (loss + ensemble_loss + 0.1 * sparse_loss + colbert_loss) / 4
3. loss += (dense_self_distill_loss + 0.1 * sparse_self_distill_loss + colbert_self_distill_loss) / 3

After :

1. return dense_scores + 0.3 * sparse_scores
2. loss = (loss + ensemble_loss + 0.1 * sparse_loss) / 3
3. loss += (dense_self_distill_loss + 0.1 * sparse_self_distill_loss) / 2

Is that valid or there is better equation ? should I reduce sparse_score ?