You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
while training, set bf16 or fp16 in TrainingArguments; while inference, set use_fp16=True in AutoModelForEmbedding or LLMRanker
The fine-tuned embedding performance during inference is worse than original?
check whether the pooling_method is correct
check whether the prompt or instruction is exactly same as training for LLM model
How can we fine-tune the BAAI/bge-m3 ColBERT model?
open-retrievals support to fine-tune the BAAI/bge-m3 colbert directly, just don't set use_fp16=True while fine-tuning, and set the learning_rate smaller
The performance is worse?
the collator and loss should be aligned, especially for triplet training with negative embeddings. The collator of open-retrievals provided is {query: value, positive: value, negative: value}. Another collator is {query: value, document: positive+negative}, if so the loss function should be treated accordingly