Skip to content

README.md

Latest commit

 

History

History
165 lines (113 loc) · 5.25 KB

README.md

File metadata and controls

165 lines (113 loc) · 5.25 KB
Raw

Deployment — YouTube labeling + DEEP-Dari dashboard

Separate package for serving DEEP-Dari, labeling real public YouTube comments, and browsing results in a bilingual EN / FA (RTL) dashboard. Training code stays under scripts/ and augmentations/.

Setup

From the repo root (after the main project requirements are installed):

pip install -r deployment/requirements.txt

Set a YouTube Data API v3 key:

# Linux / macOS / DGX
export YOUTUBE_API_KEY=your-key

# Windows PowerShell
$env:YOUTUBE_API_KEY = "your-key"

Model resolution order for the interactive /api/predict singleton (first existing checkpoint wins):

  1. outputs/ablation/A4
  2. Models/parsbert_emotion_incremental
  3. Models/parsbert_emotion

Override with --model-path or DEPLOYMENT_MODEL_PATH.

Batch jobs (New job UI / CLI) can select an ablation checkpoint by id:

model_id Primary path Automatic fallbacks
A4 (default) outputs/ablation/A4 A4/seed_42/ (multi-seed); else Models/parsbert_emotion_incremental if trained on allAug
A0 outputs/ablation/A0 —
A1 outputs/ablation/A1 Legacy outputs/full_8label_aug_* whose training_metadata.json uses Combined_Labeled_Dataset_with_fearAug.csv (highest checkpoint-*)

Also accepts config.json under checkpoint-<step>/ inside a run folder.

python -m deployment.youtube_batch --video-id VIDEO --model-id A0 --max-comments 100

GET /api/models lists availability after these lookups. Unavailable options are disabled in the dashboard. No manual symlinks required when the legacy Fear-aug or multi-seed layouts are present.

Filter-then-classify (Excluded vs Others)

Pipeline:

  1. Clean — strip URLs, @mentions, emojis; unify ي/ی and ك/ک
  2. Gate — drop comments that are empty, emoji-only, too short (< 4 tokens), or mostly non-Persian/Dari (Persian-script ratio < 0.5) → Excluded (no model call)
  3. Predict — ParsBERT on remaining text
  4. Abstain — if confidence < --min-confidence (default 0.50) → Others

So Others = model unsure on usable Dari/Persian, not emoji/Latin noise.

summary.json reports both views:

  • exclusion_rate, n_excluded
  • others_rate_among_usable, label_counts_usable
  • label_counts_all (includes Excluded)

Dashboard Run detail has a defense toggle: Usable only (default) vs All comments.

Refresh old runs (remap legacy Others+unusable_text → Excluded):

python -m deployment.reaggregate
python -m deployment.reaggregate --run-id youtube_batch_20260727T035253Z

CLI batch labeling

python -m deployment.youtube_batch --video-id VIDEO_ID --max-comments 500
python -m deployment.youtube_batch --channel-id CHANNEL_ID --max-videos 5 --max-comments 1000

Video and channel arguments accept raw IDs or full YouTube links. The CLI and New Job form normalize them the same way:

Paste Resolved to
https://www.youtube.com/watch?v=…, youtu.be/…, /shorts/…, /embed/… 11-char video id
Bare 11-char id unchanged
https://www.youtube.com/channel/UC… or bare UC… channel id
https://www.youtube.com/@Handle or @Handle UC… via YouTube API (channels.list(forHandle=…))
/c/… or /user/… best-effort resolve via API search

Examples:

python -m deployment.youtube_batch --video-id 'https://www.youtube.com/watch?v=16e75OffBTA'
python -m deployment.youtube_batch --channel-id '@SomeHandle' --max-videos 5

Invalid pastes return a clear error before the job starts.

Outputs: outputs/deployment/youtube_batch_<timestamp>/

Professional dashboard (EN / FA)

Build the React UI once (needs Node.js 18+):

cd deployment/dashboard
npm install
npm run build
cd ../..

Run API + dashboard together:

export YOUTUBE_API_KEY=your-key   # required to start jobs from the UI
uvicorn deployment.api:app --host 0.0.0.0 --port 8000

Open http://SERVER:8000

Page Features
Runs Past batch folders with KPIs
Run detail Emotion charts, filters, RTL comment table, Fear/Anger review marks
New job Start video/channel labeling; live logs
Global queue review_queue.jsonl

Language toggle EN / FA in the header (Persian uses Vazirmatn + RTL).

Dev mode (hot reload)

Terminal 1:

uvicorn deployment.api:app --reload --port 8000

Terminal 2:

cd deployment/dashboard && npm run dev

Open http://127.0.0.1:5173 (Vite proxies /api → 8000).

API (selected)

  • GET /api/health — model + YouTube key status
  • GET /api/runs — list batches
  • GET /api/runs/{id} — summary
  • GET /api/runs/{id}/comments — paginated filters
  • GET /api/runs/{id}/review / POST — review annotations
  • POST /api/jobs — start batch job
  • GET /api/jobs/{id} — job status + logs
  • POST /api/predict — single/batch text predict

Optional live chat

python -m deployment.youtube_monitor --video-id LIVE_VIDEO_ID --poll-seconds 20

Ethics

Public comments only, minimal storage, research/demo use. Alerts notify a human review queue — no auto-action.