---
license: apache-2.0
tags:
  - onnx
  - xberg
---

# Classification Models (experimental candidates)

Unmodified copies of permissively licensed upstream ONNX exports, staged by [xberg](https://github.com/xberg-io/xberg) for evaluation (xberg-io/xberg#2334, #2336). **Experimental candidates, not qualified**: no model here is an xberg default, a supported preset or a bundling decision.

| Directory | Upstream | Upstream revision | License | Task |
|---|---|---|---|---|
| `nli-minilm2-l6-h768` | [cross-encoder/nli-MiniLM2-L6-H768](https://huggingface.co/cross-encoder/nli-MiniLM2-L6-H768) | `b95119ce93d3e065de6214e38cd4a97b0f2f2c6d` | Apache-2.0 | zero-shot / NLI classification (3-way logits: 0 contradiction, 1 entailment, 2 neutral) |

Every directory keeps the upstream `LICENSE` text (`LICENSE`), the upstream model card (`UPSTREAM_README.md`), tokenizer, processor and generation configuration, and all graphs the pipeline needs. `manifest.json` lists the SHA-256 and size of every file and the upstream revision it was copied from. Consumers must pin this repository's commit revision (recorded in `crates/xberg/src/candidate_models/`) and verify the SHA-256 values; do not resolve `main`.

Attribution: each model remains the work of its upstream authors under the license listed above. Notes per model:

- `nli-minilm2-l6-h768`: RoBERTa-tokenizer MiniLMv2 cross-encoder; fp32 plus arm64 qint8 and avx2 quint8 dynamic quantizations as published upstream. Base model nreimers/MiniLMv2-L6-H768-distilled-from-RoBERTa-Large declares no license on its card (microsoft/unilm origin, MIT); the cross-encoder itself declares Apache-2.0.