electra large discriminator squad2 512

Providerahotrod
Categoryquestion-answering
LicenseApache-2.0
Downloads836.8K
Stars0

Overview

The ELECTRA-Large Discriminator, fine-tuned on SQuAD 2.0, is a specialized encoder model designed for high-precision extractive question answering. Unlike traditional BERT-based models that use masked language modeling, ELECTRA's replaced token detection pre-training allows it to learn more efficiently from the same amount of data, often resulting in superior performance on NLU benchmarks. This specific iteration supports a 512-token context window, making it suitable for analyzing medium-length documents to extract exact answer spans. For developers, this model is an ideal drop-in replacement for BERT-large in RAG pipelines or knowledge-base search tools where accuracy in identifying 'no-answer' scenarios (a key feature of SQuAD 2.0) is critical. It integrates seamlessly with the Hugging Face Transformers ecosystem, ensuring low friction for deployment in Python-based environments.

Highlights

  • Optimized for extractive QA with SQuAD 2.0 fine-tuning
  • Efficient replaced token detection for higher NLU accuracy
  • Supports context windows up to 512 tokens
  • Seamless integration via Hugging Face Transformers library
  • Apache-2.0 license for flexible commercial deployment

Usage

Install
# Install Hugging Face transformers
pip install transformers torch
SDK Usage
# Load model with transformers
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("ahotrod/electra_large_discriminator_squad2_512")
tokenizer = AutoTokenizer.from_pretrained("ahotrod/electra_large_discriminator_squad2_512")

Hugging Face Download

We recommend downloading the model via the Hugging Face CLI or Hub SDK.

Guidance：Before downloading, install huggingface_hub with:

Guidance 
pip install -U huggingface_hub

CLI Download

Download the full repository

Download the full repository 
huggingface-cli download ahotrod/electra_large_discriminator_squad2_512

Download a single file to a local folder (e.g. config.json into ./dir)

Download a single file to a local folder (e.g. config.json into ./dir) 
huggingface-cli download ahotrod/electra_large_discriminator_squad2_512 config.json --local-dir ./dir

See the official docs for more CLI options

SDK Download

SDK Download 
# 模型下载
from huggingface_hub import snapshot_download
model_dir = snapshot_download('ahotrod/electra_large_discriminator_squad2_512')

Git Download

Make sure git-lfs is installed first

Git Download 
git lfs install
git clone https://huggingface.co/ahotrod/electra_large_discriminator_squad2_512

To skip LFS large-file downloads, use:

Skip LFS 
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/ahotrod/electra_large_discriminator_squad2_512

Model files are hosted on the Hugging Face Hub — download directly via HF CLI / SDK / Git, not through this site.

PyTorch / Transformers Usage

Install Transformers

Install Transformers 
pip install -U transformers torch

Load the model and run inference

Load the model and run inference 
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained('ahotrod/electra_large_discriminator_squad2_512')
tokenizer = AutoTokenizer.from_pretrained('ahotrod/electra_large_discriminator_squad2_512')

Full Documentation

来源: HuggingFace

ELECTRA_large_discriminator language model fine-tuned on SQuAD2.0

with the following results:

code
"exact": 87.09677419354838,
  "f1": 89.98343832723452,
  "total": 11873,
  "HasAns_exact": 84.66599190283401,
  "HasAns_f1": 90.44759839056285,
  "HasAns_total": 5928,
  "NoAns_exact": 89.52060555088309,
  "NoAns_f1": 89.52060555088309,
  "NoAns_total": 5945,
  "best_exact": 87.09677419354838,
  "best_exact_thresh": 0.0,
  "best_f1": 89.98343832723432,
  "best_f1_thresh": 0.0

from script:

code
python ${EXAMPLES}/run_squad.py \
  --model_type electra \
  --model_name_or_path google/electra-large-discriminator \
  --do_train \
  --do_eval \
  --train_file ${SQUAD}/train-v2.0.json \
  --predict_file ${SQUAD}/dev-v2.0.json \
  --version_2_with_negative \
  --do_lower_case \
  --num_train_epochs 3 \
  --warmup_steps 306 \
  --weight_decay 0.01 \
  --learning_rate 3e-5 \
  --max_grad_norm 0.5 \
  --adam_epsilon 1e-6 \
  --max_seq_length 512 \
  --doc_stride 128 \
  --per_gpu_train_batch_size 8 \
  --gradient_accumulation_steps 16 \
  --per_gpu_eval_batch_size 128 \
  --fp16 \
  --fp16_opt_level O1 \
  --threads 12 \
  --logging_steps 50 \
  --save_steps 1000 \
  --overwrite_output_dir \
  --output_dir ${MODEL_PATH}

using the following system & software:

code
Transformers: 2.11.0
PyTorch: 1.5.0
TensorFlow: 2.2.0
Python: 3.8.1
OS/Platform: Linux-5.3.0-59-generic-x86_64-with-glibc2.10
CPU/GPU: Intel i9-9900K / NVIDIA Titan RTX 24GB
View official source
Join our Telegram