hari31416/indictrans2-en-indic-1B-ONNX-q4f16

IndicTrans2 1B (en→indic) — ONNX bundle [Q4F16 (4-bit Block Quantization, Lossy)]

This model is part of a suite of optimized/quantized ONNX versions of the base model. Other variants in this direction:

• FP32 (Full Precision / Base): hari31416/indictrans2-en-indic-1B-ONNX
• FP16 (Half Precision): hari31416/indictrans2-en-indic-1B-ONNX-fp16
• INT8 (Dynamic Quantization): hari31416/indictrans2-en-indic-1B-ONNX-int8
• Q4F16 (4-bit Block Quantization): hari31416/indictrans2-en-indic-1B-ONNX-q4f16 (Current)

ONNX-exported and quantized version of ai4bharat/indictrans2-en-indic-1B for in-browser and local edge inference.

• Precision: Q4F16 (4-bit Block Quantization, Lossy)
• Description: 4-bit quantization with float16 scale factors and block size of 32. Reduces model size significantly.
• Source Pipeline & Details: For pipeline details, benchmarks, and usage instructions, see the indictrans2-onnx-export GitHub repository.

Built for use with Transformers.js and onnxruntime-web in the browser, with fast BPE tokenizer.json files that don't require the SentencePiece WASM runtime.

Performance Visualizations

These charts show overall tradeoffs, language-level parity, and category breakdown.

Performance Tradeoffs & Size Comparison

Compared against the FP32 ONNX oracle on the golden evaluation fixtures.

Format	Model Size	Exact Match (Token)	Exact Match (Text)	SacreBLEU (Raw)	Latency (Mean)	Speedup vs. FP32
FP32	6.64 GB	100.00%	100.00%	100.00	69.5 ms	1.000x
FP16	3.32 GB	99.73%	99.73%	100.00	74.3 ms	0.935x
INT8	1.66 GB	89.55%	89.55%	96.27	31.4 ms	2.125x
Q4F16	850.5 MB	82.45%	82.55%	91.99	58.4 ms	1.186x

Language-Level Parity (Q4F16)

Exact match rates and translation quality (SacreBLEU / chrF) per language pair under this precision:

Language Code	Total Fixtures	Token Match Rate	Text Match Rate	SacreBLEU	SacreBLEU (chrF)
asm_Beng	50	82.0%	82.0%	92.61	97.18
ben_Beng	50	82.0%	82.0%	90.59	95.00
brx_Deva	50	84.0%	84.0%	93.64	97.53
doi_Deva	50	78.0%	78.0%	90.60	95.01
gom_Deva	50	76.0%	76.0%	87.97	95.04
guj_Gujr	50	86.0%	86.0%	94.93	98.11
hin_Deva	50	92.0%	92.0%	97.98	98.79
kan_Knda	50	82.0%	82.0%	90.60	96.27
kas_Arab	50	86.0%	86.0%	93.50	97.01
mai_Deva	50	90.0%	90.0%	92.18	97.32
mal_Mlym	50	90.0%	90.0%	95.25	99.01
mar_Deva	50	82.0%	82.0%	91.82	97.21
mni_Beng	50	82.0%	82.0%	88.86	94.03
npi_Deva	50	82.0%	82.0%	91.65	96.73
ory_Orya	50	90.0%	90.0%	95.40	98.37
pan_Guru	50	82.0%	82.0%	91.51	95.60
san_Deva	50	70.0%	70.0%	83.93	94.57
sat_Olck	50	44.0%	46.0%	76.72	86.18
snd_Arab	50	82.0%	82.0%	92.16	96.54
tam_Taml	50	82.0%	82.0%	91.81	96.57
tel_Telu	50	92.0%	92.0%	96.44	99.30
urd_Arab	50	98.0%	98.0%	99.61	99.62

Category-Level Parity (Q4F16)

Exact match rates and translation quality grouped by category types:

Category	Total Fixtures	Token Match Rate	Text Match Rate	SacreBLEU	SacreBLEU (chrF)
Generic	286	81.1%	81.1%	91.48	96.25
Lexicon	264	79.5%	79.5%	91.13	96.07
Numerals	264	84.8%	85.2%	92.92	96.73
Politics	286	84.3%	84.3%	91.88	96.28

Translation Mismatch Examples

Here is a sample of up to 5 translation mismatches compared to the FP32 oracle. Many mismatches represent minor synonym differences or spacing variations.

Mismatch #1 (Category: Politics)

• Source (eng_Latn → asm_Beng): What is the capital of India?
• Expected (FP32): ভাৰতৰ ৰাজধানী কি?
• Actual (Q4F16): ভাৰতৰ ৰাজধানী কোনখন?

Mismatch #2 (Category: Numerals)

• Source (eng_Latn → asm_Beng): Can you please help me find the nearest hospital?
• Expected (FP32): আপুনি অনুগ্ৰহ কৰি মোক নিকটতম চিকিৎসালয়খন বিচাৰি উলিওৱাত সহায় কৰিব পাৰিবনে?
• Actual (Q4F16): আপুনি অনুগ্ৰহ কৰি নিকটতম চিকিৎসালয়খন বিচাৰি উলিওৱাত মোক সহায় কৰিব পাৰিবনে?

Mismatch #3 (Category: Generic)

• Source (eng_Latn → asm_Beng): The train is delayed by two hours.
• Expected (FP32): ৰে "লখনে দুঘণ্টা পলম কৰে।
• Actual (Q4F16): ৰে'ল দুঘণ্টা পলম হয়।

Mismatch #4 (Category: Generic)

• Source (eng_Latn → asm_Beng): The crop yield has improved due to good rainfall.
• Expected (FP32): ভাল বৰষুণৰ ফলত শস্যৰ উৎপাদন উন্নত হৈছে।
• Actual (Q4F16): ভাল বৰষুণৰ ফলত শস্যৰ উৎপাদন বৃদ্ধি পাইছে।

Mismatch #5 (Category: Generic)

• Source (eng_Latn → asm_Beng): A warm cup of tea is perfect for a cold morning.
• Expected (FP32): এক গৰম কাপ চাহ ঠাণ্ডা ৰাতিপুৱাৰ বাবে উপযুক্ত।
• Actual (Q4F16): ঠাণ্ডা ৰাতিপুৱাৰ বাবে এক গৰম কাপ চাহ উপযুক্ত।

Files

• encoder_model.onnx (and optional .onnx.data weights sidecar)
• decoder_model.onnx and decoder_with_past_model.onnx (share decoder_shared.onnx.data when present)
• translate.py — self-contained Python inference helper (see Usage below)
• Fast tokenizer config files (tokenizer_src.json, tokenizer_tgt.json, tokenizer_meta.json)
• Model configuration configs (config.json, generation_config.json)

Usage Example (Python, onnxruntime)

# translate.py is included in this repo alongside the ONNX bundle.
# You can also find it (and read the full source) at:
#   https://github.com/Hari31416/indictrans2-onnx-export/blob/main/src/translate.py

from translate import IndicTransONNX

# Pass a HF repo ID for automatic download, or a local bundle directory path
model = IndicTransONNX("hari31416/indictrans2-en-indic-1B-ONNX-q4f16")
print(model.translate("Who will win the election?", src_lang="eng_Latn", tgt_lang="hin_Deva"))

Required packages:

pip install onnxruntime tokenizers huggingface-hub

License

MIT (preserved from upstream AI4Bharat).