ekinox-io/AnswerEvaluator-finetuned-int8

Answer Evaluator (fine-tuned) — ONNX INT8

Weight-only INT8 ONNX export (opset 17) of ekinox-io/AnswerEvaluator, pinned to revision 46e3b8a585d64d97f13bf1c5ec6b925cf64e60a7. Every weight matrix is INT8 with one scale per output column, embedding tables included; activations stay in FP32, so the feed-forward activation outliers that break fully dynamic quantization cannot degrade the output. The model evaluates French student answers with four independent binary labels.

Files

• model.int8.onnx — weight-only INT8 model (112.6 MB)
• config.json — model configuration and label mapping
• tokenizer files — text preprocessing
• quantization_report.json — quantized operator counts, size reduction and FP32 logit parity
• artifact_manifest.json — files managed by this export pipeline

Inputs and output

• input_ids: int64 [batch, sequence]
• attention_mask: int64 [batch, sequence]
• logits: float32 [batch, 4]

Labels, in output order: is_true, is_relevant, is_complete, is_syntax_correct. Apply sigmoid independently to each logit; do not use argmax.

The expected text format is:

Question : ... [SEP] Réponse attendue : ... [SEP] Réponse élève : ... [SEP] Texte : ...

Usage

This repository is gated. Request access, then authenticate with hf auth login.

import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download
from transformers import AutoConfig, AutoTokenizer

repo_id = "ekinox-io/AnswerEvaluator-finetuned-int8"
text = (
    "Question : Que boit l'enfant ? [SEP] "
    "Réponse attendue : L'enfant boit du lait. [SEP] "
    "Réponse élève : Il boit du lait. [SEP] "
    "Texte : Après le repas, l'enfant boit un verre de lait."
)

tokenizer = AutoTokenizer.from_pretrained(repo_id, token=True)
config = AutoConfig.from_pretrained(repo_id, token=True)
inputs = tokenizer(text, return_tensors="np", truncation=True, max_length=512)
model_path = hf_hub_download(repo_id, "model.int8.onnx", token=True)
session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
logits = session.run(
    ["logits"],
    {
        "input_ids": inputs["input_ids"].astype(np.int64),
        "attention_mask": inputs["attention_mask"].astype(np.int64),
    },
)[0]
probabilities = 1 / (1 + np.exp(-logits[0]))
labels = [config.id2label[index] for index in range(config.num_labels)]
predictions = dict(zip(labels, (probabilities >= 0.5).tolist()))
print(predictions)