npc0/qwen3-embedding-rerank-projection-v2

Qwen3-Embedding Rerank Projection

Metric-learning MLP projection that maps Qwen/Qwen3-Embedding-0.6B embeddings (1024-dim) to a compact L2-normalized 256-dim space for fast dot-product reranking.

Supersedes npc0/qwen3-embedding-rerank-projection.

Model details

• Base model: Qwen/Qwen3-Embedding-0.6B
• Architecture: Linear(1024, 512) → LayerNorm → ReLU → Dropout(0.1) → Linear(512, 256) → LayerNorm → ReLU → Dropout(0.1) → Linear(256, 256) → LayerNorm, with L2 normalization on the output.
• Loss used for this checkpoint: smooth_kernel_v2
• Quantization scale (calibration): 432.4123025768911

Metrics

Metric	Value
beir_nq_mrr@10	0.4019
beir_nq_ndcg@10	0.438
beir_nq_recall@100	0.7209
best_epoch	8
epochs_trained	15
loss_eps	0.3
loss_sigma	0.1
loss_temperature	0.05
v1_ndcg@10	0.4354

Usage

import torch
from safetensors.torch import load_file
from run_train import MLPProjection

model = MLPProjection(input_dim=1024, output_dim=256)
model.load_state_dict(load_file("model.safetensors"))
model.eval()

def rerank(query_emb, doc_embs):
    # Forward already L2-normalizes, so the dot product is cosine similarity.
    with torch.no_grad():
        q = model(query_emb)
        d = model(doc_embs)
    scores = (q * d).sum(dim=-1)
    return scores.argsort(descending=True)

Limitations

• Trained with metric learning; zero-shot generalization to other corpora is not guaranteed.
• The projection is specific to the 1024-dim embeddings of Qwen/Qwen3-Embedding-0.6B; do not feed other embedders' vectors.

Citation

If you used this in your research, please cite:

@misc{xu2026qwen3rerank,
  title        = {Qwen3-Embedding-0.6B Soft-Label Rerank Projection},
  author       = {Yuan Xu},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/npc0/qwen3-embedding-rerank-projection-v2}},
  note         = {Soft-label (kernel-smoothed) projection head on frozen Qwen3-Embedding-0.6B}
}