You will fine-tune a compact DistilBERT encoder to decide whether a news article is fake or real, export the model to an 8-bit ONNX file, and run it fully client-side in a single-page React / Vite app.
The popular Fake and Real News corpus (~44 k articles) fits on a laptop, and INT8 quantisation shrinks the model to ≈ 18 MB, keeping browser latency below 200 ms with WebGPU.

Build a browser-only fake-news detector that takes an article (headline + text) and returns

• a probability (0 = real, 1 = fake)
• a quick verdict badge

All inference must remain in the user’s tab—no server calls.

Misinformation detection is topical, involves longer inputs than tweets, and requires more than surface-level cues—perfect for demonstrating practical NLP skills without huge compute.

1. Public Git repo with a concise README.md describing
   • dataset provenance & licence link
   • training hyper-parameters
   • evaluation metrics (F1, accuracy)
2. model.onnx ≤ 25 MB and tokenizer.json committed under /public/model/.
3. Browser demo at <your-url>/fakenews/ featuring
   • textarea (or drop-zone) for article text
   • probability bar ± verdict badge
   • console log of inference latency
4. Short screencast (< 2 min) or live URL proving the app works offline (reload with Wi-Fi disabled).

conda create -n fakenews python=3.10
conda activate fakenews
pip install "transformers>=4.40" datasets evaluate accelerate \
           "optimum[onnxruntime,gpu]" onnxruntime-web==1.17.0 \
           @huggingface/transformers      # JS side

ONNX Runtime Web ≥ 1.17 adds official WebGPU support.

from datasets import load_dataset
# The Kaggle CSV can be loaded directly once downloaded locally
news = load_dataset(
    "csv",
    data_files={"train": "train.csv", "test": "test.csv"}
)

Split 10 % of training into validation; truncate/clean text as desired.

• Task: sequence classification
• Epochs: 2
• Learning rate: 2 e-5
• max_length: 384
• Batch size: 8–16 (GPU-dependent)
  Monitor F1 on the validation set.

Keep the best checkpoint (save_total_limit=1) and log metrics with evaluate.

optimum-cli export onnx --model path/to/best onnx/ --quantize dynamic

Expect an ~18 MB INT8 model.

import { pipeline } from "@huggingface/transformers";

const clf = await pipeline(
  "text-classification",
  "/public/model",
  { quantized: true }
);
const { label, score } = (await clf(articleText))[0];

import * as ort from "onnxruntime-web";

const session = await ort.InferenceSession.create("/model.onnx", {
  executionProviders: ["webgpu", "wasm"]
});

Good luck—help us separate facts from fiction!