PlanTL-GOB-ES/roberta-base-bne-sqac and other spanish models

[iagovar]

Model description

Hi, I tried to export PlanTL-GOB-ES/roberta-base-bne-sqac to ONNX with the conversion script in the repo, but although the model works with transformers + pytorch in python, I always get this error when using it in JS:

RangeError: Invalid array length
    at Function._call (/home/iagovar/MEGA/MEGAsync/web/mywebsite.com/repo/calendar/node_modules/@xenova/transformers/src/tokenizers.js:2528:78)
    at Function.closure [as tokenizer] (/home/iagovar/MEGA/MEGAsync/web/mywebsite.com/repo/calendar/node_modules/@xenova/transformers/src/utils/core.js:62:28)
    at Function._call (/home/iagovar/MEGA/MEGAsync/web/mywebsite.com/repo/calendar/node_modules/@xenova/transformers/src/pipelines.js:356:27)
    at closure (/home/iagovar/MEGA/MEGAsync/web/mywebsite.com/repo/calendar/node_modules/@xenova/transformers/src/utils/core.js:62:28)
    at main (/home/iagovar/MEGA/MEGAsync/web/mywebsite.com/repo/calendar/sandbox/ai.js:35:22)
    at process.processTicksAndRejections (node:internal/process/task_queues:95:5) {stack: 'RangeError: Invalid array length
    at Funct…ions (node:internal/process/task_queues:95:5)', message: 'Invalid array length'}

In theory, given the lists you provided, this model should be no problem.

Prerequisites

• [X] The model is supported in Transformers (i.e., listed here)
• [X] The model can be exported to ONNX with Optimum (i.e., listed here)

Additional information

No response

Your contribution

I don't think so, I'm a very junior coder.

[xenova]

Hi there 👋 Can you share the code you used that produces this error? Also, have you possibly uploaded the models to the HF Hub? If so, I can do some testing myself.

[iagovar]

Hi Xenova, first of all, thanks for your work! Without you it would be much harder.

Here's the pastebin of my code, just a test: https://pastebin.com/1GBxePEP

Here's the HF repo with the model converted to onnx using the script you guys provided: https://huggingface.co/iagovar/roberta-base-bne-sqac-onnx

[xenova]

Thanks for the additional context. Unfortunately, I am unable to reproduce the issue. Which version of transformers.js are you using? Running the following code:

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

const answerer = await pipeline('question-answering', 'roberta-base-bne-sqac', { quantized: false })

const question = "¿Dónde se celebra el concierto?";
const context = "El concierto de juanes se celebrará en el palacio de la ópera, a las 8 de la tarde";
const output = await answerer(question, context);
console.log(output);

produces:

{ answer: ' palacio de la ópera', score: 0.5542509822824329 }

Also, could you update the repo structure of this model so that the files in this folder are in the main directory? https://huggingface.co/iagovar/roberta-base-bne-sqac-onnx/tree/main/roberta-base-bne-sqac

[iagovar]

Hmmm, could the node version affect? I'm using 18+. Also, I'm using let { pipeline, env } = await import('@xenova/transformers'); as I'm using commonJS through the application I want to embbed this in. I wonder if this might have something to do with it.

And yes, I will update the repo, I uploaded it in a hurry through web interface, sorry.

[xenova]

Odd indeed, as neither of those should be an issue. Would you mind putting together a minimal repository/application for me to inspect in closer detail?

[iagovar]

Hi Xenova, a bit late, but here it is. I was using Transformers.js 2.12.0 according to its package.json

I did a little test script, for better readability here's the pastebin: https://pastebin.com/VxibUaW9

This script uses two different QA Models. Both fail. Here's the output: https://pastebin.com/0VPTJBfZ

The same input works perfectly fine in python: https://pastebin.com/mCF3vUX0

---

I noticed that if cut the cleanContext variable in about half the size, the first model does make the inference, while the second still fails.

The problem is that I can't really do this, as the intention is to use QA models to precisely extract information from messy dumps of code, with the intention of avoiding to use APIs (I will loop through thousands of records from a database).

My intuition is that there's some data structure that overflows in some way. Can't say really, just speculating, because it works fine in python.

I translated the models to ONNX using the provided script. Here are the files, in case it helps: https://files.fm/u/db4mfn8p94