Scoreboar Twitter/X Virality

Scoreboar is the original local-first Chrome MV3 extension that adds compact scoring labels to X/Twitter timeline posts and lightweight hints while drafting a post. This public package uses the same extension scripts, DOM detection, popup, icons, styles, and build pipeline as the working private extension; the only publishing-specific addition is a Hugging Face download script for the large ONNX model assets.

What it does

• Adds one compact badge per detected article[data-testid="tweet"] on https://x.com/* and https://twitter.com/*.
• Adds debounced composer hints for X post textareas/contenteditable composer boxes.
• Composer scoring passively includes attached-media state plus current-account handle/follower/following/post/verified metadata when X has already exposed it in the page DOM or loaded state.
• Passively reads author metadata only when X has already loaded it into same-page GraphQL responses.
• Runs local ONNX inference in a Chrome offscreen document.
• Keeps model, tokenizer, ONNX Runtime Web, and WASM assets packaged locally under dist/.

It has no backend, no telemetry, no auth handling, no cloud sync, no extra X API/profile probing, and no runtime CDN/model fetches.

Install in Chrome

npm install
npm run build:hf

Then:

1. Open chrome://extensions.
2. Enable Developer mode.
3. Click Load unpacked.
4. Select this repo’s dist/ folder.
5. Open https://x.com or https://twitter.com.

Hugging Face model download

Large model files are not committed to git. npm run build:hf downloads the reference model assets from Hugging Face before running the original extension build.

Default model repo:

siimh/scoreboar-twitter-x-virality

Useful commands:

npm run download:model
npm run build:hf

Optional pinning:

SCOREBOAR_HF_REPO=siimh/scoreboar-twitter-x-virality \
SCOREBOAR_HF_REVISION=<commit-sha-or-tag> \
npm run build:hf

For private or gated repos, set HF_TOKEN or HUGGING_FACE_HUB_TOKEN. Never commit tokens.

The download script writes the files into the exact paths expected by the original build:

artifacts/model/v5-full.onnx
model/v5-source/tokenizer/tokenizer.json

The original build then packages them as:

dist/extension/assets/model/v5-full.onnx
dist/extension/assets/tokenizer/tokenizer.json

The stable runtime filename is v5-full.onnx; this file contains the final/latest validated v7-lineage export.

API example

The repo also includes a tiny Express service in examples/express-service/ for people who want to run the ONNX model behind their own API instead of inside Chrome. It uses onnxruntime-node, the same tokenizer, and the same metadata preprocessing as the extension.

npm run download:model
cd examples/express-service
npm install
npm run build
npm start

Then score text with:

curl -s http://localhost:8787/score \
  -H 'content-type: application/json' \
  -d '{"text":"I built a tiny local model that tells you when your tweet is probably dead.","metadata":{"hasMedia":false}}'

This is only an example wrapper. The Chrome extension still runs locally and does not call this API.

Development

npm run typecheck
npm test
npm run build:hf
npm run assert:dist
npm run assert:manifest
npm run assert:no-remote-assets

Model and data summary

• Runtime artifact: v5-full.onnx stable filename containing the final/latest validated v7-lineage model export.
• Base encoder: answerdotai/ModernBERT-base.
• Architecture: shared ModernBERT text encoder + 12-field metadata fusion + feature heads + 5-way ordinal outcome head.
• Approximate training corpus: ~60K Twitter/X.com posts total.
  • ~50K viral/high-engagement posts.
  • ~10K random/baseline posts.
  • Refreshed with recent posts from roughly the last 18 months.
• Grok/teacher enrichment targets: 20 total auxiliary targets: 12 numeric scores, 5 boolean flags, and 3 categorical labels.
• Runtime/browser ONNX exposes the 5-way outcome head plus 12 numeric and 5 boolean feature heads.
• Validation snapshot: 58.73% exact 5-bucket accuracy and 98.34% within ±1 bucket.

Enriched shape, in plain terms:

• Inputs at runtime: post text plus 12 metadata values: has_media, created_at_hour_sin, created_at_hour_cos, created_at_day_sin, created_at_day_cos, log_author_followers, log_author_following, log_author_tweets, author_verified, hashtag_count, mention_count, url_count.
• Main output: 5 performance buckets: very_low, low, medium, high, very_high.
• Auxiliary numeric outputs: virality_score, hook_quality, clarity_score, novelty_score, emotional_intensity, controversy_level, shareability_score, conversation_potential, authenticity_score, urgency_level, call_to_action_strength, trend_alignment.
• Auxiliary boolean outputs: is_rage_bait, is_clickbait, is_ai_slop, needs_context, has_clear_takeaway.
• Categorical labels used during training: primary_emotion, target_audience, content_type. These were training supervision; the browser runtime does not need to show them.

See MODEL_CARD.md for the full model card.

Project layout

manifest.config.ts             source manifest used by original build
extension/                     original MV3 entrypoints, popup, icons, page listener
src/                           original DOM detection, scoring UI, guardrails, runtime contracts
scripts/build-extension.mjs    original extension build script
scripts/download-hf-assets.mjs Hugging Face model/tokenizer downloader
examples/express-service/      optional Node.js API wrapper for the ONNX model
fixtures/                      local X-like fixture pages for tests
tests/                         original unit/integration tests
MODEL_CARD.md                  Hugging Face model documentation

Guardrails

• Runtime/model assets are packaged locally under dist/extension/assets/.
• The extension does not load scripts, WASM, tokenizers, or model files from remote URLs at runtime.
• The extension does not make X API requests.
• Scoreboar may passively parse already-loaded same-page X GraphQL responses for author metadata.
• The score is directional. Use ranges/buckets such as medium–high, not exact truth claims.