trumancai/Revela-500M

Model Summary

Revela-500M is a self-supervised bi-encoder dense-retriever trained with the Revela objective on raw Wikipedia text.
It uses the 500 M-parameter Qwen 2.5-0.5B backbone and was trained on 320 K Wikipedia batches (batch size = 16).
The in-batch attention mechanism enables fully self-supervised learning without manually-mined relevance labels.
See the paper for full details.

• Repository: TRUMANCFY/Revela
• Training Dataset: trumancai/revela_training_corpus

Other Links

Binary	Description
trumancai/Revela-1b	1 B-parameter variant (LLaMA-3.2-1B backbone).
trumancai/Revela-500M	← current repo
trumancai/Revela-135M	135 M-parameter variant (SmolLM2-135M backbone).
trumancai/Revela-code-1b	1 B-parameter code-retriever.
trumancai/Revela-code-500M	500 M-parameter code-retriever.
trumancai/Revela-code-135M	135 M-parameter code-retriever.
trumancai/revela_training_corpus	Wikipedia training corpus.
trumancai/revela_code_training_corpus	Code training corpus.

Usage

Evaluate with the customised mteb fork:

from mteb.model_meta import ModelMeta
from mteb.models.repllama_models import RepLLaMAWrapper, _loader
import mteb, torch

revela_qwen_500m = ModelMeta(
    loader=_loader(
        RepLLaMAWrapper,
        base_model_name_or_path="Qwen/Qwen2.5-0.5B",
        peft_model_name_or_path="trumancai/Revela-500M",
        device_map="auto",
        torch_dtype=torch.bfloat16,
    ),
    name="trumancai/Revela-500M",
    languages=["eng_Latn"],
    open_source=True,
    revision="2206071b5fe41cdae695dd705b4ccc6afc63f759",
    release_date="2025-04-13",
)
model = revela_qwen_500m.loader()

mteb.MTEB(tasks=["SciFact", "NFCorpus"]).run(model=model, output_folder="results/Revela-500M")

Licence

Citation