Original model: https://huggingface.co/BAAI/bge-reranker-v2-m3 Original authors: the BAAI / FlagEmbedding team (Chen et al., 2024), built on the multilingual BAAI/bge-m3 backbone (XLM-RoBERTa large) ONNX export used: https://huggingface.co/onnx-community/bge-reranker-v2-m3-ONNX (fp32 onnx/model.onnx + onnx/model.onnx_data, produced by the onnx-community/convert-to-onnx space — the BAAI repository ships no ONNX) Converted by: Lucie666, using burn-onnx — format only

This is not an original model, and no part of it is my work. It is a mechanical format conversion of BAAI/bge-reranker-v2-m3 — nothing was trained, fine-tuned, distilled, quantised or modified. No new weights were produced. All credit belongs to the authors above.

If you are looking for the model itself, go to BAAI/bge-reranker-v2-m3. This repository only exists so people running Burn don't each redo the conversion.

Note on metadata: Hugging Face's base_model_relation field only accepts adapter, merge, quantized or finetune. None describes a pure format conversion, so the field is deliberately left unset rather than filled with an inaccurate value — these weights are not quantised, they are the original f32 values.

The file model.bpk holds the same weights as the upstream ONNX export, re-serialised into Burn's burnpack format so they can be loaded by a pure-Rust inference stack — no Python, no PyTorch, no ONNX Runtime at inference time.

Why this exists

bge-reranker-v2-m3 is the strongest openly licensed multilingual cross-encoder of its generation (about 100 languages, long inputs). At 2.2 GB and 24 layers it is a server model; in rag3weaver it is the quality-first reranker behind SearchOptions.rerank, next to the 470 MB mmarco-mMiniLMv2-L12-H384-v1 (multilingual default) and the 90 MB ms-marco-MiniLM-L-6-v2 (English, browser).

Provenance

BAAI/bge-reranker-v2-m3   model.safetensors
        │
        │  onnx-community/convert-to-onnx   (PyTorch 2.8 export → onnx/model.onnx + model.onnx_data, 2 271 088 656 bytes)
        ▼
        │  burn-onnx 0.22.0-pre.1   (mechanical ONNX → Burn conversion, LoadStrategy::Bytes)
        ▼
model.bpk        weights, burnpack format
model.rs         model graph, generated Rust source (not distributed here)

Nothing in this pipeline is hand-written. config.json and tokenizer.json used at conversion and at runtime are the ones from the BAAI repository (the onnx-community copy only differs by _name_or_path and the transformers version).

A caveat before you regenerate. Burnpack serialisation is not byte-deterministic: two builds from the same ONNX produce files of identical size but different bytes. The tensor values are unaffected. The checksum below verifies this download, not a reproduction.

Reproducing it

rust

// build.rs
use burn_onnx::{ModelGen, LoadStrategy};

fn main() {
    ModelGen::new()
        .input("onnx/model.onnx")       // model.onnx_data must sit next to it
        .out_dir("model/")
        .load_strategy(LoadStrategy::Bytes)
        .run_from_script();
}

The weights exceed the 2 GB ONNX protobuf limit and live in external data; burn-onnx ≥ 0.22.0-pre.1 reads them without any setting (0.21 panics). The generated graph uses burn::nn::LinearLayout::Col for the classification head, which exists from burn 0.22.0-pre.2 — generated code and runtime must match versions. Conversion took 28 s wall and 2.4 GB of RAM; loading the model for inference takes about 6.7 GB.

Verifying the download

sha256  3ed858274ab4661332058318c8b961f0ac822af4aed899187557745107fb32e3
size    2271128324 bytes  (2.12 GiB)

Model interface

rust

pub fn forward(
    &self,
    input_ids: Tensor<2, Int>,
    attention_mask: Tensor<2, Int>,
) -> Tensor<2>    // logits [B, 1]

The graph is the full XLMRobertaForSequenceClassification: <s> token → dense (Linear(1024→1024) + tanh) → out_proj Linear(1024→1). The output is the raw logit: higher means more relevant; apply a sigmoid yourself if you want a probability. Only the order is meaningful.

Tokenisation is not included: use the upstream tokenizer.json from BAAI/bge-reranker-v2-m3 (SentencePiece Unigram, 250 002 entries; <s> = 0, <pad> = 1, </s> = 2). It shares its vocabulary with bge-m3 but not its normalizer (a trailing-whitespace strip), so do not substitute one for the other. There are no token_type_ids: a (query, passage) pair is one sequence <s> query </s></s> passage </s>, padded with id 1. The model accepts up to 8192 positions; rerankers are normally run at 512.

Numerical parity

Checked against a reference candle implementation (XLM-RoBERTa backbone from candle-transformers + the upstream classification head — classifier.dense (tanh) and classifier.out_proj — loaded from model.safetensors, CPU) on seven (query, passage) pairs: the Berlin example of the MS MARCO model card in English, the same triple in French, and a cross-language pair (French query, English passage):

pair                                                        burn (wgpu)    candle (CPU)      |Δ|
------------------------------------------------------------------------------------------------
EN berlin population / 3.5 million inhabitants                 6.798565        6.798559     5.7e-06
EN berlin population / Metropolitan Museum of Art            -11.028526      -11.028533     6.7e-06
EN berlin population / Berlin Wall fell in 1989               -9.748907       -9.748901     5.7e-06
FR combien … berlin / 3,5 millions d'habitants                 6.165921        6.165920     4.8e-07
FR combien … berlin / Metropolitan Museum of Art             -11.032015      -11.032012     2.9e-06
FR combien … berlin / mur de Berlin tombé en 1989            -10.509789      -10.509794     5.7e-06
FR query / EN "3.5 million inhabitants" passage                5.617373        5.617373     9.5e-07

max |Δ| = 6.7e-6 on raw logits, rankings identical in both languages and across languages. The residual is f32 accumulation noise from a different operation order.

Runtime backend for this check: Burn + wgpu/Vulkan on an AMD Radeon AI PRO R9700 (Navi 48, RDNA4, gfx1201) via RADV.

License

Apache-2.0, inherited from BAAI/bge-reranker-v2-m3. All credit for the model belongs to its authors:

bibtex

@misc{li2023making,
  title={Making Large Language Models A Better Foundation For Dense Retrieval},
  author={Chaofan Li and Zheng Liu and Shitao Xiao and Yingxia Shao},
  year={2023},
  eprint={2312.15503},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

@misc{chen2024bge,
  title={BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation},
  author={Jianlv Chen and Shitao Xiao and Peitian Zhang and Kun Luo and Defu Lian and Zheng Liu},
  year={2024},
  eprint={2402.03216},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

Lucie666/bge-reranker-v2-m3-burnpack

bge-reranker-v2-m3-burnpack