MIRA-IndicLID

MIRA-IndicLID is a state-of-the-art, lightweight Language Identification (LID) model specialized for the 22 Constitutional Languages of India.

This model introduces a significantly more balanced training approach, leveraging a mixture of formal, social, and synthetic data to achieve superior generalization across diverse web domains and scripts (Native, Romanized, and Code-Mixed).

Model Details

• Architecture: FastText (Binary + Quantized)
• Size: ~902 MB (.ftz format)
• Supported Languages: 22 Indic Languages
• Key Feature: High robustness on Romanized Indic content and Code-Mixed (Hinglish) text.

Supported Languages

The model supports the following 22 languages: Assamese (as), Bengali (bn), Bodo (brx), Dogri (doi), Gujarati (gu), Hindi (hi), Kannada (kn), Kashmiri (ks), Konkani (kok), Maithili (mai), Malayalam (ml), Manipuri (mni), Marathi (mr), Nepali (ne), Oriya (or), Punjabi (pa), Sanskrit (sa), Santali (sat), Sindhi (sd), Tamil (ta), Telugu (te), Urdu (ur).

Label Mapping

Language	ISO Code	fastText Label
Assamese	as	__label__as
Bengali	bn	__label__bn
Bodo	brx	__label__brx
Dogri	doi	__label__doi
Gujarati	gu	__label__gu
Hindi	hi	__label__hi
Kannada	kn	__label__kn
Kashmiri	ks	__label__ks
Konkani	kok	__label__kok
Maithili	mai	__label__mai
Malayalam	ml	__label__ml
Manipuri	mni	__label__mni
Marathi	mr	__label__mr
Nepali	ne	__label__ne
Oriya	or	__label__or
Punjabi	pa	__label__pa
Sanskrit	sa	__label__sa
Santali	sat	__label__sat
Sindhi	sd	__label__sd
Tamil	ta	__label__ta
Telugu	te	__label__te
Urdu	ur	__label__ur

Datasets Used

Combination of data from:

• IndicCorp: ai4bharat/IndicCorpV2
• Sangraha: ai4bharat/sangraha
• IndicSentiment: ai4bharat/IndicSentiment
• GlotLID: cis-lmu/glotlid-corpus
• FLORES: facebook/flores
• Hinglish: findnitai/english-to-hinglish

The training data includes both native scripts and Romanized (transliterated) text, with temperature-based sampling (T=0.3) to handle low-resource languages and script balancing.

Training Data & Strategy

The model was trained on a comprehensive mixture of ~2.8 billion characters (balanced sampling) from the following sources:

• Formal Native Corpus: Clean text from IndicCorp v2 and FLORES-200.
• Social & Noisy Web: Real-world data from IndicSentiment and custom social media crawls.
• Large-Scale Web Crawl: Sangraha dataset, stratified by script and length.
• Synthetic Hard Pairs: Adversarial samples generated via Groq Synthetic Data to resolve common language confusions (e.g., Hindi vs. Marathi in Devanagari).
• Romanized & Code-Mixed: Extensive transliterated and Hinglish data to ensure script-agnostic identification.

Benchmarking Metrics (Accuracy & Macro-F1)

The following tables summarize the performance of MIRA-IndicLID compared to several state-of-the-art baselines.

Accuracy (%)

Dataset	MIRA-IndicLID	AI4B-IndicLID	GlotLID	FastText	OpenLID
Test-Set	96.86	-	84.63	69.29	68.32
FLORES	98.32	99.39	99.69	100.00	99.08
Sentiment	96.75	99.30	100.00	74.88	74.88
Sangraha	87.81	89.00	82.77	68.81	68.35
IndicLID-Full-Test	97.48	96.75	95.89	90.95	89.77
Comprehensive-Native	98.02	97.39	96.66	89.78	88.11
Comprehensive-Roman	81.82	45.39	34.60	0.13	0.00
Comprehensive-Mixed	100.00	14.29	52.38	0.00	0.00
Entire-Comprehensive	88.16	65.01	58.47	34.40	33.68
Groq-Synthetic	70.69	64.23	64.92	39.20	38.16

Macro-F1 (%)

Dataset	MIRA-IndicLID	AI4B-IndicLID	GlotLID	FastText	OpenLID
Test-Set	95.15	-	74.80	60.84	59.27
FLORES	99.64	99.91	99.94	100.00	99.85
Sentiment	98.76	99.84	100.00	89.91	88.78
Sangraha	74.57	76.49	76.19	65.21	64.21
IndicLID-Full-Test	97.08	92.59	94.07	76.77	75.32
Comprehensive-Native	97.60	95.69	94.87	78.55	76.68
Comprehensive-Roman	79.95	40.19	34.03	0.16	0.00
Comprehensive-Mixed	100.00	25.00	68.75	0.00	0.00
Entire-Comprehensive	90.08	64.68	68.40	45.91	46.28
Groq-Synthetic	72.50	66.12	71.56	49.86	49.16

Timing Summary (Inference Time in Seconds)

Lower is better.

Dataset	MIRA-IndicLID	AI4B-IndicLID	GlotLID	FastText	OpenLID
Test-Set	31.7	-	242.5	68.4	66.2
FLORES	0.1	0.0	0.4	0.2	0.2
Sentiment	0.2	0.1	1.0	0.3	0.3
Sangraha	9.6	237.9	117.0	24.8	23.6
IndicLID-Full-Test	8.2	4.8	36.5	18.2	17.7
Comp-Native	0.2	0.1	0.8	0.5	0.5
Comp-Roman	0.3	229.6	1.0	0.5	0.4
Comp-Mixed	0.0	1.1	0.0	0.0	0.0
Entire-Comprehensive	0.3	231.9	1.7	0.7	0.6
Groq-Synthetic	0.2	32.5	2.2	0.5	0.4

AI4B-IndicLID Comparison

The results show that while AI4B-IndicLID performs exceptionally well on standard datasets like FLORES and Sentiment, MIRA-IndicLID significantly outperforms it on Romanized and Mixed script tasks (Comp-Roman, Comp-Mixed, Groq-Synthetic).

Usage

Prerequisites

Install the required libraries:

pip install fasttext huggingface_hub

Inference Code

You can use the model with the fasttext library.

import fasttext
import os

# Path to our model binary
OUR_MODEL_PATH = "mira_indic_lid.ftz"  # Update with your local path
fasttext.FastText.eprint = lambda x: None # Suppress warnings

# Load Model
model = fasttext.load_model(OUR_MODEL_PATH)

# Experiment 1: Native Script
text = " किमिदं भवति!!!  😂 🤨🤨 सर्वं विचित्रम् दृश्यu"
labels, probs = model.predict(text, k=5, threshold=0.1) 
print(f"Prediction: {labels}, Probability: {probs}")

# Experiment 2: Romanized Text
text = " angni nwngni khobor ma? 😂😂 gwsw jwng nwngkhou!! !!"
labels, probs = model.predict(text, k=5, threshold=0.0) 
print("Text:", text)

# Print label and probability pairs
for l, p in zip(labels, probs):
    print(f"{l.replace('__label__',''):5s}  {p:.4f}")

Parameters for .predict():

1. k: (int) Number of labels to return.
2. threshold: (float) Minimum probability to return a label.

Authors

Developed by Team Project MIRA