A multimodal embedding model based on Qwen3-VL for text and image retrieval tasks. Built with Tevatron framework and fine-tuned using LoRA.

• Multimodal retrieval: Encode text queries and images into a shared embedding space
• Based on Qwen3-VL-4B-Instruct: Leverages powerful vision-language foundation model
• Efficient training: Uses LoRA for parameter-efficient fine-tuning
• MTEB evaluation: Compatible with MTEB benchmark tasks

# Install with uv
uv sync

Train a retriever model using the provided training script:

bash train.sh

The training uses:

• DeepSpeed for distributed training
• LoRA fine-tuning of Qwen3-VL-4B-Instruct
• Contrastive learning with temperature scaling
• EOS pooling with normalized embeddings

Configure your training data in dataset_config.yaml.

With MTEB wrapper: Check inference_mteb.py

With Transformers: Check inference_transformers.py

import mteb

model_meta = mteb.get_model_meta('eagerworks/eager-embed-v1')
model = model_meta.load_model()

# Get benchmarks and extract tasks from them
benchmarks = mteb.get_benchmarks(["ViDoRe(v2)"])
tasks = []
for benchmark in benchmarks:
    tasks.extend(benchmark.tasks)
print(tasks)
# Run evaluation with reduced batch size to save CUDA memory
results = mteb.evaluate(model=model, tasks=tasks, encode_kwargs={"batch_size": 8})

print("Evaluation complete!")
print(results)

For more examples, see evaluate_mteb.py

python merge_lora_and_push.py \
    --adapter_path run2_8x5090 \
    --push_to_hub_id eagerworks/eager-embed-v1 \
    --dtype float32

• Base Model: Qwen3-VL-4B-Instruct (~3B parameters)
• Embedding Dimension: 2560
• Pooling: Last token (EOS) pooling
• Normalization: L2 normalized embeddings
• Training: Contrastive learning with in-batch negatives

Apache 2.0

@article{EagerEmbed,
  title={Eager Embed V1: Multimodal Dense Embeddings for Retrieval},
  author={Juan Pablo Balarini},
  year={2025},
  publisher={Eagerworks},
  url={https://github.com/eagerworks/eager-embed}
}