Original model: https://huggingface.co/cross-encoder/ms-marco-MiniLM-L-6-v2 Original authors: Nils Reimers and the sentence-transformers team (cross-encoder series, trained on MS MARCO Passage Ranking), fine-tuned from Microsoft's MiniLM (Wang et al., 2020) 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 cross-encoder/ms-marco-MiniLM-L-6-v2 — 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 cross-encoder/ms-marco-MiniLM-L-6-v2. 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/model.onnx, 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

A cross-encoder scores a (query, passage) pair directly and is the standard way to re-rank the top candidates of a retrieval stage. At ~90 MB this one is small enough to run in a browser (WebGPU) or an embedded Rust process, which is what it does in rag3weaver as the reranker behind SearchOptions.rerank.

Provenance

cross-encoder/ms-marco-MiniLM-L-6-v2   onnx/model.onnx   (91 011 230 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.

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")       // from cross-encoder/ms-marco-MiniLM-L-6-v2
        .out_dir("model/")
        .load_strategy(LoadStrategy::Bytes)
        .run_from_script();
}

The generated graph uses burn::nn::LinearLayout::Col for the pooler and classifier (the head was exported Gemm-style), which exists from burn 0.22.0-pre.2 — generated code and runtime must match versions.

Verifying the download

sha256  f2c416115ca43604b18a4e7da3c0651ea0cdb10994f1e6a60f19185304d9acd6
size    90883844 bytes  (86.7 MiB)

Model interface

rust

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

The graph is the full BertForSequenceClassification: CLS token → pooler (Linear(384→384) + tanh) → classifier Linear(384→1). The output is the raw logit (the upstream default activation is the identity): 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 (BERT uncased WordPiece, [PAD] = 0, [CLS] = 101, [SEP] = 102, max 512 positions). Inputs are pairs (query, passage) encoded as one sequence [CLS] query [SEP] passage [SEP] with token_type_ids 0 for the query segment and 1 for the passage — the upstream tokenizer produces them for a pair.

Numerical parity

Checked against a reference candle implementation (candle-transformers BertModel

the upstream bert.pooler.dense (tanh) and classifier weights loaded from model.safetensors, CPU) on five (query, passage) pairs, including the model card's Berlin example:

pair                                              burn (wgpu)    candle (CPU)      |Δ|
------------------------------------------------------------------------------------
berlin population / "3.5 million inhabitants"       8.648815       8.648816     9.5e-07
berlin population / "Metropolitan Museum of Art"  -11.352621     -11.352626     4.8e-06
berlin population / "Berlin Wall fell in 1989"     -9.009550      -9.009547     2.9e-06
rust borrow checker / matching passage              4.504333       4.504328     5.3e-06
rust borrow checker / cake recipe                 -11.164450     -11.164444     5.7e-06

max |Δ| = 5.7e-6 on raw logits, rankings identical; the candle values match the numbers published on the upstream model card. The residual is f32 accumulation noise across 6 layers 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 cross-encoder/ms-marco-MiniLM-L-6-v2. All credit for the model belongs to its authors:

bibtex

@inproceedings{reimers-2019-sentence-bert,
  title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
  author = "Reimers, Nils and Gurevych, Iryna",
  booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
  year = "2019",
  url = "https://arxiv.org/abs/1908.10084"
}

@article{wang2020minilm,
  title={MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers},
  author={Wang, Wenhui and Wei, Furu and Dong, Li and Bao, Hangbo and Yang, Nan and Zhou, Ming},
  journal={arXiv preprint arXiv:2002.10957},
  year={2020}
}

Lucie666/ms-marco-minilm-l6-v2-burnpack

ms-marco-minilm-l6-v2-burnpack