MahaVani

MahaVani is a professionally organized collection of 36 PEFT/LoRA translation adapters for 9 Indian tribal languages across Hindi and English directions.

The companion complete-model repository is MahaVani_Purna.

🚀 Deployment & Serving

Ready-to-use deployment guide and code are in deployment/:

• deployment/DEPLOYMENT.md — full hosting guide (prompt format, full-vs-adapter hosting, FastAPI, Modal serverless, vLLM/TGI, OpenAI-compatible integration, quantization).
• deployment/translator.py — reusable Translator (loads full models or base + hot-swappable LoRA adapters; greedy decoding; batching).
• deployment/api_service.py — FastAPI REST service exposing all 36 directions.
• deployment/modal_service.py — serverless GPU HTTPS endpoint (scales to zero).

Prompt: system = "Translate the text below to {TARGET_LANGUAGE}." + user = <source text>. For → tribal directions, {TARGET_LANGUAGE} is the tribal language name (e.g. "Bhili", "Santhali", "Betta Kurumba"). Decode greedily (do_sample=False) for deterministic translation.

Structure

MahaVani/
├── metrics.csv
├── metrics.json
├── model_index.json
├── <tribal-language>/<source>_to_<target>/
│   ├── README.md
│   ├── adapter/
│   ├── metrics/
│   ├── logs/
│   └── training_state/
└── source/

Architecture and Method

• Adaptation method: LoRA / PEFT adapters
• Precision: bf16
• Training: 2 epochs per direction
• Objective: target-only translation loss
• Evaluation: held-out test split; no test leakage

Metrics

#	Direction	Language	chrF++	BLEU	Test rows
0	bettakuruba_to_hindi	bettakuruba	59.35	42.13	423
1	hindi_to_bettakuruba	bettakuruba	57.24	23.83	423
2	bettakuruba_to_english	bettakuruba	69.80	50.20	419
3	english_to_bettakuruba	bettakuruba	69.44	37.72	419
4	bhili_to_hindi	bhili	75.14	60.81	2000
5	hindi_to_bhili	bhili	51.77	25.52	2000
6	bhili_to_english	bhili	48.36	21.23	2000
7	english_to_bhili	bhili	36.36	13.50	2000
8	garo_to_hindi	garo	55.90	36.51	2000
9	hindi_to_garo	garo	46.96	17.04	2000
10	garo_to_english	garo	68.08	54.84	2000
11	english_to_garo	garo	50.73	21.25	2000
12	gondi_to_hindi	gondi	68.20	53.01	2000
13	hindi_to_gondi	gondi	47.80	22.97	2000
14	gondi_to_english	gondi	50.01	24.99	2000
15	english_to_gondi	gondi	37.10	14.33	2000
16	kokborok_to_hindi	kokborok	53.51	34.12	829
17	hindi_to_kokborok	kokborok	51.46	21.09	829
18	kokborok_to_english	kokborok	67.53	56.37	797
19	english_to_kokborok	kokborok	57.59	26.48	797
20	koya_to_hindi	koya	43.53	23.24	271
21	hindi_to_koya	koya	31.93	4.82	271
22	koya_to_english	koya	54.87	38.81	272
23	english_to_koya	koya	33.50	6.00	272
24	kui_to_hindi	kui	36.23	14.25	1843
25	hindi_to_kui	kui	31.35	4.67	1843
26	kui_to_english	kui	40.23	15.68	1843
27	english_to_kui	kui	30.01	4.22	1843
28	mundari_to_hindi	mundari	72.82	59.85	2000
29	hindi_to_mundari	mundari	49.95	24.13	2000
30	mundari_to_english	mundari	36.44	16.63	2000
31	english_to_mundari	mundari	24.98	7.04	2000
32	santhali_to_hindi	santhali	71.94	57.34	2000
33	hindi_to_santhali	santhali	73.04	56.51	2000
34	santhali_to_english	santhali	53.21	27.26	2000
35	english_to_santhali	santhali	59.07	42.63	2000

Load an adapter

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model = "sarvamai/sarvam-translate"
repo_id = "adivaanihf/MahaVani"
subfolder = "bhili/bhili_to_hindi/adapter"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, repo_id, subfolder=subfolder)
model.eval()

Continue fine-tuning

Use the selected direction adapter/ folder as the starting PEFT checkpoint. Keep the same base model, tokenizer, prompt format, and target-only loss masking. See each direction's training_state/ and metrics/ folders.

Limitations

Metrics are automatic MT metrics; human evaluation is recommended before production use. Raw training data is not included.