mteb_llama_nv_embed_reasoning_3b.py

# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0.
"""
MTEB encoder and ModelMeta for nvidia/llama-nv-embed-reasoning-3b.
"""

from mteb.models.model_meta import ModelMeta
from mteb.models.model_implementations.nvidia_models import (
    LlamaEmbedNemotron,
    LlamaEmbedNemotron_CITATION,
    llama_embed_nemotron_evaluated_languages,
    llama_embed_nemotron_training_datasets,
)
from mteb.types import PromptType

BRIGHT_TASK_INSTRUCTIONS = {
    "BrightBiologyRetrieval": "Given a Biology post, retrieve relevant passages.",
    "BrightEarthScienceRetrieval": "Given an Earth Science post, retrieve relevant passages.",
    "BrightEconomicsRetrieval": "Given an Economics post, retrieve relevant passages.",
    "BrightPsychologyRetrieval": "Given a Psychology post, retrieve relevant passages.",
    "BrightRoboticsRetrieval": "Given a Robotics post, retrieve relevant passages.",
    "BrightStackoverflowRetrieval": "Given a Stack Overflow post, retrieve relevant passages.",
    "BrightSustainableLivingRetrieval": "Given a Sustainable Living post, retrieve relevant passages.",
    "BrightLeetcodeRetrieval": "Given a Coding problem, retrieve relevant passages.",
    "BrightPonyRetrieval": "Given a Pony question, retrieve relevant passages.",
    "BrightAopsRetrieval": "Given a Math problem, retrieve relevant passages.",
    "BrightTheoremQAQuestionsRetrieval": "Given a Math problem, retrieve relevant passages.",
    "BrightTheoremQATheoremsRetrieval": "Given a Math problem, retrieve relevant passages.",
}

BRIGHT_PASSAGE_PREFIX = "passage: "

class LlamaNvEmbedReasoning(LlamaEmbedNemotron):
    """LlamaNvEmbedReasoning for reasoning with BRIGHT benchmark prompts."""

    def __init__(self, model_name: str, revision: str, device: str | None = None, **kwargs) -> None:
        super().__init__(model_name, revision=revision, device=device)
        self.max_seq_length = kwargs.get("max_seq_length", 8192)

    def _get_base_instruction(self, task_metadata, prompt_type: PromptType | None) -> str:
        task_name = task_metadata.name
        if task_name in BRIGHT_TASK_INSTRUCTIONS:
            if prompt_type == PromptType.document:
                return ""
            return BRIGHT_TASK_INSTRUCTIONS[task_name]
        return super()._get_base_instruction(task_metadata, prompt_type)

    def encode(
        self,
        inputs,
        *,
        task_metadata,
        hf_split: str = "",
        hf_subset: str = "",
        prompt_type: PromptType | None = None,
        **kwargs,
    ):
        task_name = task_metadata.name
        if task_name in BRIGHT_TASK_INSTRUCTIONS and prompt_type == PromptType.document:
            prefix = BRIGHT_PASSAGE_PREFIX
        else:
            instruction = self._get_task_specific_instruction(task_metadata, prompt_type)
            prefix = self.format_instruction(instruction, prompt_type)
        return self._extract_embeddings(inputs, instruction=prefix, **kwargs)

LLAMA_NV_EMBED_REASONING_3B_META = ModelMeta(
    loader=LlamaNvEmbedReasoning,
    loader_kwargs=dict(max_seq_length=8192),
    name="nvidia/llama-nv-embed-reasoning-3b",
    model_type=["dense"],
    languages=llama_embed_nemotron_evaluated_languages,
    open_weights=True,
    revision="main",
    release_date="2026-02-23", 
    n_parameters=3_212_749_824,
    memory_usage_mb=6000,
    embed_dim=3072,
    license="https://huggingface.co/nvidia/llama-nv-embed-reasoning-3b/blob/main/LICENSE",
    max_tokens=8192,
    reference="https://huggingface.co/nvidia/llama-nv-embed-reasoning-3b",
    similarity_fn_name="cosine",
    framework=["PyTorch", "Transformers"],
    use_instructions=True,
    training_datasets=llama_embed_nemotron_training_datasets,
    public_training_code=None,
    public_training_data=None,
    citation=LlamaEmbedNemotron_CITATION,
)