You can deploy this model and start earning money today!

Kritika00/
hinglish-ner-xlmr

Name: hinglish-ner-xlmr
Availability: InStock

Views

No views yet

🇮🇳 Hinglish NER — Named Entity Recognition for Hindi-English Mixed Text

The first production-grade open-source NER model for Hinglish (Hindi-English code-mixed) text.

What is Hinglish?

Hinglish is Hindi-English code-mixed text — how 600M+ Indians actually write online.

"Virat ne Mumbai mein Zomato se biryani order ki aur 500 rupaye pay kiye."
→ Virat (PERSON), Mumbai (LOCATION), 500 rupaye (MONEY)

No standardised spelling. No existing NER dataset. This project fills that gap.

Model Performance

Entity	Precision	Recall	F1
DATE	1.000	1.000	1.000
ORG	0.786	0.688	0.733
LOCATION	0.667	0.444	0.533
PERSON	0.514	0.439	0.474
PRODUCT	0.000	0.000	0.000
MONEY	—	—	— (not in fresh test set)
Overall	0.614	0.479	0.538

Evaluated on a clean held-out test set (fresh sentences never seen during training).
Dev F1 during training: 74.10 (step 2,400, P=79.12, R=69.68)
Baseline (Gliner zero-shot, no fine-tuning): F1 = 59.8

Known limitations: PRODUCT generalises poorly (only 115 training examples). PERSON recall is low due to high spelling variation in Hinglish names.

Tech Stack

Tool	Role
spaCy 3.7+	NER training framework
XLM-RoBERTa	Multilingual encoder (pre-trained on 100 languages)
Gliner	Zero-shot baseline + pre-annotation
Label Studio	Manual annotation
FastAPI	REST API deployment
Gradio	Interactive demo UI
HuggingFace Hub	Model + dataset publishing
DVC	Data version control

Project Structure

hinglish-ner/
├── data/
│   ├── raw/                        ← 2,169 collected Hinglish sentences
│   ├── annotated/annotations.json  ← Label Studio export
│   └── processed/                  ← train/dev/test .spacy files
├── models/
│   └── hinglish-ner-v1/model-best/ ← trained spaCy model
├── app/
│   └── main.py                     ← FastAPI server
├── scripts/
│   ├── baseline_gliner.py          ← zero-shot baseline
│   ├── convert_to_spacy.py         ← Label Studio → DocBin
│   └── baseline_evaluate.py        ← evaluation script
├── demo.py                         ← Gradio demo
├── config.cfg                      ← spaCy training config
└── requirements.txt

Quickstart

1. Clone and set up environment

bash

git clone https://github.com/kritikachoudhary/hinglish-ner.git
cd hinglish-ner
python -m venv hinglish-env
source hinglish-env/bin/activate
pip install -r requirements.txt

2. Download the model

bash

# From HuggingFace
python -c "
from huggingface_hub import snapshot_download
snapshot_download('kritikachoudhary/hinglish-ner-xlmr', local_dir='models/hinglish-ner-v1/model-best')
"

3. Run FastAPI server

bash

uvicorn app.main:app --reload
# → http://localhost:8000
# → http://localhost:8000/docs  (Swagger UI)

4. Run Gradio demo

bash

python demo.py
# → http://localhost:7860

API Reference

`POST /ner` — Single text

bash

curl -X POST http://localhost:8000/ner \
  -H "Content-Type: application/json" \
  -d '{"text": "Virat ne Mumbai mein Zomato se biryani order ki"}'

Response:

json

{
  "text": "Virat ne Mumbai mein Zomato se biryani order ki",
  "entities": [
    {"text": "Virat",  "label": "PERSON",   "start": 0,  "end": 5},
    {"text": "Mumbai", "label": "LOCATION", "start": 9,  "end": 15}
  ]
}

`POST /ner/batch` — Multiple texts (max 100)

bash

curl -X POST http://localhost:8000/ner/batch \
  -H "Content-Type: application/json" \
  -d '{"texts": ["Priya Delhi gayi.", "Shah Rukh Khan Netflix pe hai."]}'

`GET /` — Health check

bash

curl http://localhost:8000/
# {"status": "ok", "model": "...", "labels": ["DATE", "LOCATION", ...]}

Entity Types

Label	Examples
PERSON	Virat, Shah Rukh Khan, Priya
ORG	Google, Zomato, Netflix, Amazon
LOCATION	Mumbai, Delhi, Bangalore
PRODUCT	iPhone 15, biryani
DATE	next Friday, kal, agle hafte
MONEY	500 rupaye, 80000 mein

Training Details

Base model: XLM-RoBERTa (trained on 100 languages including Hindi + English)
Framework: spaCy 3.7 + spacy-transformers
Dataset: 1,000+ manually annotated Hinglish sentences (YouTube comments, product reviews)
Annotation tool: Label Studio with Gliner pre-annotation
Training: Kaggle T4 GPU, 3,000 steps, LR = 4e-5
Best checkpoint: step 2,400 (F1 = 74.10)

Limitations

MONEY entity boundaries can be noisy (e.g. includes surrounding words)
Rare names and novel ORGs may be missed
Performance degrades on very short texts (< 5 words)
Not tested on formal Hindi (Devanagari script) — designed for romanised Hinglish only

Kritika00/hinglish-ner-xlmr

🇮🇳 Hinglish NER — Named Entity Recognition for Hindi-English Mixed Text

What is Hinglish?

Model Performance

Tech Stack

Project Structure

Quickstart

1. Clone and set up environment

2. Download the model

3. Run FastAPI server

4. Run Gradio demo

API Reference

POST /ner — Single text

POST /ner/batch — Multiple texts (max 100)

GET / — Health check

Entity Types

Training Details

Limitations

Kritika00/
hinglish-ner-xlmr

`POST /ner` — Single text

`POST /ner/batch` — Multiple texts (max 100)

`GET /` — Health check