Implementation:FlagOpen FlagEmbedding FlagAutoModel From Finetuned Validation

Source
FlagEmbedding/inference/auto_embedder.py:L23-111

Summary

This documents using FlagAutoModel.from_finetuned() specifically for loading fine-tuned checkpoints for validation. The underlying API is the same as FlagAutoModel.from_finetuned() used for inference, but applied in a validation context.

Key difference from inference loading: model_name_or_path points to the local output_dir (the fine-tuning output directory), and model_class MUST be explicitly specified.

Signature

Same as FlagAutoModel.from_finetuned().

Import

from FlagEmbedding import FlagAutoModel

Example

model = FlagAutoModel.from_finetuned(
    model_name_or_path="./fine_tuned_model",
    model_class="encoder-only-base"
)
embeddings = model.encode(["test query", "test passage"])
similarity = embeddings[0] @ embeddings[1]
print(f"Similarity: {similarity:.4f}")

Related Pages

• Principle:FlagOpen_FlagEmbedding_Finetuned_Embedder_Validation