Dhee-Vagmi-Core

A multilingual machine-translation model covering English and 12 Indian languages, trained from scratch in the M2M100 architecture. It translates directly between any supported pair (not only to and from English).

Version: v2 · Languages: English, Hindi, Bengali, Marathi, Tamil, Telugu, Kannada, Malayalam, Odia, Gujarati, Punjabi, Assamese, Urdu.

What this model is for

Sentence and short-passage translation across English and 12 Indian languages, in any direction. It is not a general-purpose LLM or a document summariser.

What you get

• A fully open, from-scratch multilingual MT checkpoint -- no proprietary base model in the lineage, so the only upstream term is the training data's licence.
• Direct any-to-any translation across 13 languages from a single model, loadable with the standard transformers M2M100ForConditionalGeneration class -- no custom modelling code.
• A shipped SentencePiece tokenizer (spm.model) with FLORES-style language tokens, so you can reproduce the exact tokenization the model trained with.

Usage

from transformers import M2M100ForConditionalGeneration
import sentencepiece as spm, torch

model = M2M100ForConditionalGeneration.from_pretrained("dheeyantra/dhee-vagmi-core")
sp = spm.SentencePieceProcessor(model_file="spm.model")   # shipped in this repo
FLORES = {"en": "eng_Latn", "hi": "hin_Deva", "bn": "ben_Beng", "mr": "mar_Deva",
          "ta": "tam_Taml", "te": "tel_Telu", "kn": "kan_Knda", "ml": "mal_Mlym",
          "or": "ory_Orya", "gu": "guj_Gujr", "pa": "pan_Guru", "as": "asm_Beng",
          "ur": "urd_Arab"}

def translate(text, src, tgt):
    ids = [sp.piece_to_id(FLORES[src])] + sp.encode(text)[:126] + [3]
    out = model.generate(torch.tensor([ids]), max_new_tokens=128,
                         forced_bos_token_id=sp.piece_to_id(FLORES[tgt]))
    keep = [t for t in out[0].tolist() if t not in (0, 2, 3)
            and t not in {sp.piece_to_id(v) for v in FLORES.values()}]
    return sp.decode(keep)

print(translate("How are you?", "en", "hi"))

Evaluation

metric	value
n	1000
BLEU	8.22
chrF++	24.48

Limitations

• Best on short-to-medium sentences; very long inputs are truncated at encode time.
• No guarantee of terminology consistency or safe handling of code-mixed input.

Files in this repository

• README.md
• config.json
• generation_config.json
• model.safetensors
• spm.model

Licence

Released under CC-BY-NC-4.0 — free to use, share and adapt for non-commercial purposes with attribution.

Commercial use requires a separate licence — contact contact@dheeyantra.com.

This is a non-commercial release regardless of any upstream permissive label; see the attribution below for the full licence lineage.

Attribution

• Base model: none — trained from scratch (M2M100 architecture).
• Data: AI4Bharat/Samanantar (CC-BY-4.0); evaluated on FLORES-200.
• Released non-commercially by choice; contact us for commercial terms.

Citation

@misc{dhee_vagmi_core_v2,
  title  = {Dhee-Vagmi-Core},
  author = {DheeYantra Research Labs},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/dheeyantra/dhee-vagmi-core}},
  note   = {Non-commercial release; contact contact@dheeyantra.com for commercial use}
}