hari31416/indictrans2-indic-indic-dist-320M-ONNX-fp16

IndicTrans2 200M (indic→indic) — ONNX bundle [FP16 (Half Precision, Lossless)]

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-indic-indic-dist-320M-ONNX
• FP16 (Half Precision): hari31416/indictrans2-indic-indic-dist-320M-ONNX-fp16 (Current)
• INT8 (Dynamic Quantization): hari31416/indictrans2-indic-indic-dist-320M-ONNX-int8
• Q4F16 (4-bit Block Quantization): hari31416/indictrans2-indic-indic-dist-320M-ONNX-q4f16

ONNX-exported and quantized version of ai4bharat/indictrans2-indic-indic-dist-320M for in-browser and local edge inference.

• Precision: FP16 (Half Precision, Lossless)
• Description: Converts weights and activations to float16. Lossless tier, recommended for Apple Silicon (MPS) and CUDA runtimes.
• 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	1.87 GB	100.00%	100.00%	100.00	23.0 ms	1.000x
FP16	980.2 MB	99.82%	99.82%	100.00	27.4 ms	0.840x
INT8	497.1 MB	72.18%	72.36%	87.13	16.5 ms	1.475x
Q4F16	711.6 MB	45.91%	46.36%	71.64	28.3 ms	0.829x

Language-Level Parity (FP16)

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	100.0%	100.0%	100.00	100.00
ben_Beng	50	100.0%	100.0%	100.00	100.00
brx_Deva	50	100.0%	100.0%	100.00	100.00
doi_Deva	50	100.0%	100.0%	100.00	100.00
gom_Deva	50	100.0%	100.0%	100.00	100.00
guj_Gujr	50	100.0%	100.0%	100.00	100.00
hin_Deva	50	100.0%	100.0%	100.00	100.00
kan_Knda	50	100.0%	100.0%	100.00	100.00
kas_Arab	50	100.0%	100.0%	100.00	100.00
mai_Deva	50	100.0%	100.0%	100.00	100.00
mal_Mlym	50	100.0%	100.0%	100.00	100.00
mar_Deva	50	100.0%	100.0%	100.00	100.00
mni_Beng	50	100.0%	100.0%	100.00	100.00
npi_Deva	50	100.0%	100.0%	100.00	100.00
ory_Orya	50	98.0%	98.0%	99.06	99.74
pan_Guru	50	100.0%	100.0%	100.00	100.00
san_Deva	50	100.0%	100.0%	100.00	100.00
sat_Olck	50	98.0%	98.0%	99.35	99.57
snd_Arab	50	100.0%	100.0%	100.00	100.00
tam_Taml	50	100.0%	100.0%	100.00	100.00
tel_Telu	50	100.0%	100.0%	100.00	100.00
urd_Arab	50	100.0%	100.0%	100.00	100.00

Category-Level Parity (FP16)

Exact match rates and translation quality grouped by category types:

Category	Total Fixtures	Token Match Rate	Text Match Rate	SacreBLEU	SacreBLEU (chrF)
Generic	286	99.7%	99.7%	99.86	99.95
Lexicon	264	100.0%	100.0%	100.00	100.00
Numerals	264	99.6%	99.6%	99.80	99.90
Politics	286	100.0%	100.0%	100.00	100.00

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: Generic)

• Source (npi_Deva → ory_Orya): नयाँ नीति आगामी महिनादेखि लागु हुनेछ।
• Expected (FP32): ଆଗାମୀ ମାସରୁ ନୂଆ ନୀତି କାର୍ଯ୍ଯ଼କାରୀ ହେବ।
• Actual (FP16): ନୂଆ ନୀତି ଆଗାମୀ ମାସରୁ କାର୍ଯ୍ଯ଼କାରୀ ହେବ।

Mismatch #2 (Category: Numerals)

• Source (san_Deva → sat_Olck): एतत् चलच्चित्रं सत्यकथायाः आधारेण निर्मितम् आसीत्।
• Expected (FP32): ᱪᱚᱞᱚᱛ ᱪᱤᱛᱟᱹᱨ ᱫᱚ ᱢᱤᱫᱴᱟᱝ ᱥᱟᱡᱟᱹᱭᱟᱱ ᱠᱟᱹᱦᱱᱤ ᱪᱮᱛᱟᱱ ᱴᱮᱦᱟᱴ ᱠᱟᱛᱮ ᱵᱮᱱᱟᱣ ᱟᱠᱟᱱᱟ ᱾
• Actual (FP16): ᱪᱚᱞᱚᱛ ᱪᱤᱛᱟᱹᱨ ᱫᱚ ᱢᱤᱫᱴᱟᱝ ᱥᱟᱡᱟᱹᱭᱟᱱ ᱠᱟᱹᱦᱱᱤ ᱪᱮᱛᱟᱱ ᱴᱮᱦᱟᱴ ᱛᱟᱦᱮᱸ ᱠᱟᱱᱟ ᱾

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-indic-indic-dist-320M-ONNX-fp16")
print(model.translate("चुनाव कौन जीतेगा?", src_lang="hin_Deva", tgt_lang="tam_Taml"))

Required packages:

pip install onnxruntime tokenizers huggingface-hub

License

MIT (preserved from upstream AI4Bharat).