textcnn-essay-model

TextCNN Essay Scorer

Лёгкая CNN-модель для автоматической оценки эссе.

Архитектура

• Embedding dim: 128
• Filters: 128 × [2, 3, 4, 5]
• Tokenizer: bert-base-uncased

Загрузка

import json
import torch
from huggingface_hub import snapshot_download
from modeling_textcnn import TextCNNRegressor

# Скачиваем репо
path = snapshot_download(repo_id="rezka00/textcnn-essay-model")

# Читаем конфиг
with open(f"{path}/textcnn_config.json") as f:
    cfg = json.load(f)

# Восстанавливаем модель
model = TextCNNRegressor(
    vocab_size=cfg["vocab_size"],
    embed_dim=cfg["embed_dim"],
    num_filters=cfg["num_filters"],
    filter_sizes=cfg["filter_sizes"],
    dropout=cfg["dropout"],
)

# Загружаем веса
from safetensors.torch import load_file
state_dict = load_file(f"{path}/model.safetensors")
model.load_state_dict(state_dict)
model.eval()