Modelpublic

atjsh/llmlingua-2-js-xlm-roberta-large-meetingbank-onnx-v4

sourceHugging Facemitupdated 1mo agoView on Hugging Face

0likes561downloads

Model Card

LLMLingua-2 XLM-R ONNX v4 conversion archive

This is a neutral, reproducible format-conversion archive of `microsoft/llmlingua-2-xlm-roberta-large-meetingbank` at source revision ebaba9b0e874dadd3003ffcff828e4397e568089.

The source repository declares MIT. The retained license and conversion notice are in LICENSE and NOTICE. These files were not retrained.

Provenance

—Converter reference: onnx-community/convert-to-onnx@af0ef3070cef5e863b371628ad3dcdc10de4d091
—Export: optimum-cli export onnx --model <pinned-local-source> --task token-classification <output-directory>
—Variants: python conversion/quantize.py onnx/model.onnx
—Resolved environment: conversion/requirements.txt
—Converter output: conversion/export.log and conversion/quantize.log
—Machine-readable file sizes and SHA-256 hashes: conversion/manifest.json

The export command exited successfully and reported a maximum reference/ONNX logit difference of 7.2479248046875e-05, above its 1e-05 validation tolerance. The warning is preserved in conversion/export.log and does not alter this archive's converter-trust publication rule.

The q8 dtype maps to onnx/model_quantized.onnx, an exact copy of the official recipe's signed int8 output.

Artifacts

Dtype	File	Bytes	SHA-256
fp32 graph	`onnx/model.onnx`	488,518	`2cddfd767604e06699123e2640178ef3231699b6362fda364fa5152ff7baae4e`
fp32 external data	`onnx/model.onnx_data`	2,235,785,232	`a957df2691421facdac43eb4df22598980345f2e8c7da5a467a91e18b674ce63`
fp16	`onnx/model_fp16.onnx`	2	`7648eb81c4e9fef7b51879c6fcd6fb4b8863eea0cb2a978c508711dc52c80041`
int8	`onnx/model_int8.onnx`	560,745,894	`e2ee14513c0527822c3701f4adbae16e3e95463cb05d7058cbef79ab3b396e31`
uint8	`onnx/model_uint8.onnx`	560,745,966	`716baafeff90f70a94e945c7824e5c0456508ba8a744d30c07dbb3b997ef02ec`
q8	`onnx/model_quantized.onnx`	560,745,894	`e2ee14513c0527822c3701f4adbae16e3e95463cb05d7058cbef79ab3b396e31`
q4	`onnx/model_q4.onnx`	1,217,047,880	`cb14ee1a4145f7388aae32130ff528c498507a455dcb19b3206639f8893d7a57`
q4f16	`onnx/model_q4f16.onnx`	684,410,603	`7ce164c4ac2b5f9d0b06d52a364efaedaff05b2224816bd8134b71e7d0ec17a7`
bnb4	`onnx/model_bnb4.onnx`	1,217,047,880	`cb14ee1a4145f7388aae32130ff528c498507a455dcb19b3206639f8893d7a57`

Observed runtime result

conversion/runtime-cpu.json records a three-run ONNX Runtime 1.28.0 CPU smoke on an Apple M4. fp32, int8, uint8, q8, q4, and bnb4 loaded, returned finite [1, 13, 2] logits, and repeated byte-identically. fp16 failed because its two-byte emitted graph has no opset; q4f16 failed with an ONNX type mismatch. These are observations from that run, not a runtime-support, numerical-fidelity, or product-suitability claim.