HienDuong/nemotron-embed-stage1-2048-deepedu

Exported Embedding Model

This model package was exported from the Stage 2 training run:

• Run directory: 20260717_004322_stage2_nemotron_stage1_2048
• Export date: 2026-08-28

It is intended to be downloaded directly from Hugging Face and tested immediately.

Model Info

• Model: HienDuong/nemotron-embed-stage1-2048-deepedu
• Model link: https://huggingface.co/HienDuong/nemotron-embed-stage1-2048-deepedu
• Base HF model: nvidia/llama-nemotron-embed-1b-v2
• Base link: https://huggingface.co/nvidia/llama-nemotron-embed-1b-v2

Prompt Behavior

• Query prefix: query:
• Document prefix: passage:

Loading

from sentence_transformers import SentenceTransformer

model = SentenceTransformer(
    "HienDuong/nemotron-embed-stage1-2048-deepedu",
    token="HF_READ_TOKEN_IF_PRIVATE",
    trust_remote_code=True,
)

Best Checkpoint Selection

• load_best_model_at_end: True
• metric_for_best_model: eval_DeepEdu_dot_ndcg@10
• best_metric: 0.5825205475182422
• best_model_checkpoint: output/20260717_004322_stage2_nemotron_stage1_2048/checkpoints/checkpoint-200

Artifact Layout

• root: loadable Sentence Transformers model
• artifacts/stage2_manifest.json: training/export manifest
• artifacts/best_summary.json: best-checkpoint summary
• artifacts/post_eval_summary.json: post-train benchmark summary
• artifacts/post_eval/: raw post-eval outputs

Notes

• trust_remote_code=True is required.
• The base model uses custom code.
• For retrieval, keep the exact prefixes:
  • query: query:
  • document: passage:
• If the downstream team does not use the correct prefixes, retrieval quality may drop.
• The base model supports longer context and dynamic embedding dimensions, so downstream evaluation should be done with the target serving setup.