clarenceorg/llmlingua-2-onnx

LLMLingua-2 (ONNX, int8)

ONNX conversion of microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank, for running prompt compression in the browser with Transformers.js.

Only the format was changed: PyTorch → ONNX, then dynamic int8 quantization. The weights are not retrained or otherwise modified.

Why this exists

The upstream repository ships PyTorch weights only, and browsers can only run ONNX. Existing community conversions were unusable for redistribution because they declared no licence — and with no licence, the default is "all rights reserved". This conversion is made from the Apache 2.0 original so that the same permissions carry over, attribution included.

Files

file	size
onnx/model_quantized.onnx	179 MB — int8, this is what you want

The fp32 export is not published: it is 710 MB and no runtime here needs it. Regenerate it from the upstream weights if required.

Fidelity

The export was verified against the original PyTorch model rather than assumed correct. Numbers are the mean/max absolute difference in p(keep) per token, and the number of flipped decisions (keep vs. drop) over a French/English probe set:

	mean	max	flipped
this export, fp32	0.00000	0.0000	0
this export, int8 per-channel	0.01217	0.1049	1
int8 per-tensor (not used)	0.02916	0.2154	1

The fp32 figure is the one that matters: the export itself is exact, and the remaining gap comes only from quantization.

per_channel=True is deliberate — per-tensor quantization degrades roughly 3× more for 1 MB less. reduce_range=True was also measured and made things clearly worse, so it is not used.

Verified to load and run under onnxruntime-web 1.14 (the version bundled by Transformers.js v2) in a real browser.

Usage

import { pipeline } from '@xenova/transformers';

const pipe = await pipeline(
  'token-classification',
  'clarenceorg/llmlingua-2-onnx',
  { quantized: true }
);

The model has no id2label, so labels come out as LABEL_0 / LABEL_1. LABEL_1 means "keep the token"; p(keep) for LABEL_0 is 1 - score.

Two practical notes, both learned the hard way:

• The window is 512 positions, not 512 words — beyond that the pipeline truncates silently. Batch your input.
• The pipeline omits some tokens from its output (its index field skips values). Aligning a cursor over that stream desynchronises it permanently. Re-tokenize yourself and join scores back by index.

Citation

@inproceedings{pan2024llmlingua2,
  title     = {{LLMLingua-2}: Data Distillation for Efficient and Faithful
               Task-Agnostic Prompt Compression},
  author    = {Pan, Zhuoshi and Wu, Qianhui and Jiang, Huiqiang and Xia, Menglin
               and Luo, Xufang and Zhang, Jue and Lin, Qingwei and Rühle, Victor
               and Yang, Yuqing and Lin, Chin-Yew and Zhao, H. Vicky and Qiu, Lili
               and Zhang, Dongmei},
  booktitle = {Findings of ACL 2024},
  year      = {2024}
}

Licence

Apache 2.0, inherited from the upstream model. Copyright Microsoft Corporation. See the NOTICE file in this repository.

Converted for Clarence, a client-side PII anonymizer for LLM prompts.