A conversational/casual-register domain adaptation of ai4bharat/indictrans2-en-indic-1B, produced with experience replay (mixed fine-tuning) + model souping (WiSE-FT-style weight averaging with the base, α=0.6).

Scope of the claim — please read. On the chrF2 metric, this model beats IndicTrans2-1B on conversational test sets across all 21 Indic languages (mean +6.2, paired-bootstrap significant, p ≤ 0.004 on the hardest subtitle tests) while matching the base on general-domain FLORES (mean change −0.17, all within ±0.7 chrF). This is a chrF result, not a verified human-quality improvement: part of the conversational gain plausibly reflects style-matching to casual subtitle references rather than better adequacy, and a controlled human evaluation has not yet confirmed a perceived quality gain. Use accordingly. See the paper for the full, honest evaluation and limitations.

Results (chrF2, base → this model)

Lang	conv	FLORES
Hindi	53.9 → 56.8 (+2.9)	60.0 → 59.6
Tamil	52.7 → 55.1 (+2.4)	64.1 → 64.4
Bengali	68.7 → 73.6 (+4.8)	58.1 → 58.1
Malayalam	45.7 → 48.4 (+2.7)	61.1 → 61.4
Marathi	70.2 → 77.3 (+7.1)	55.0 → 54.5
mean (21 langs)	+6.2	−0.17

Conversational gains are largest where the test set is BPCC-H-Daily (in-distribution style); the most reliable gains are on the harder subtitle tests (Hindi/Tamil/Malayalam). See paper §6–§7.

Intended use & limitations

Use for: English → Indic translation of casual / conversational text (chat, dialogue, subtitles).
Direction: English → Indic only. Indic → English / Indic ↔ Indic are untested.
Register: tuned toward casual register; for formal/news text the base model is at least as good.
Metric caveat: improvements are measured by chrF2 and are partly stylistic; not human-verified.
Low-resource ceiling: very low-resource languages (e.g. Santali, Sanskrit, Kashmiri) remain bounded by the base model's absolute quality.

How to use

python

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from IndicTransToolkit import IndicProcessor   # pip install IndicTransToolkit

REPO = "adipras1407/indictrans2-en-indic-1B-conversational"
# Load the tokenizer from the base checkpoint (recommended).
tok = AutoTokenizer.from_pretrained("ai4bharat/indictrans2-en-indic-1B", trust_remote_code=True)
model = AutoModelForSeq2SeqLM.from_pretrained(REPO, trust_remote_code=True).eval()
ip = IndicProcessor(inference=True)

src = ["Are you coming over tonight?", "I missed you so much, yaar."]
batch = ip.preprocess_batch(src, src_lang="eng_Latn", tgt_lang="hin_Deva")
enc = tok(batch, return_tensors="pt", padding=True, truncation=True, max_length=256)
with torch.inference_mode():
    out = model.generate(**enc, num_beams=5, max_length=256)
hyp = ip.postprocess_batch(tok.batch_decode(out, skip_special_tokens=True), lang="hin_Deva")
print(hyp)

Use the appropriate target tag (e.g. tam_Taml, ben_Beng, mar_Deva, …) for other languages.

Training

Base: ai4bharat/indictrans2-en-indic-1B (1.1B params).
Recipe: mixed fine-tuning on 538k pairs (284k conversational + 254k BPCC-H-Wiki general anchor), 1 epoch, bf16, Adafactor, single A10G GPU; then weight-averaged with the base at α=0.6.
Conversational data: OPUS OpenSubtitles, BPCC-H-Daily, Tatoeba (English→Indic, per-pair target-language tag), filtered for length/script; Hindi additionally LaBSE-filtered ≥0.70.

Data licensing & attribution

This is a derivative of IndicTrans2 (© AI4Bharat, MIT license) — please cite their work. Training data: BPCC (AI4Bharat, open license), Tatoeba (CC-BY 2.0 FR), and OpenSubtitles via OPUS (research use; consult OPUS/OpenSubtitles terms before any commercial use). The released weights are provided for research; verify the upstream data terms for your use case.

Citation

bibtex

@misc{singh2026convindic,
  title  = {Conversational Domain Adaptation of IndicTrans2 across 21 Indic Languages
            via Experience Replay and Model Soups},
  author = {Aditya Pratap Singh},
  year   = {2026},
  eprint = {arXiv:XXXX.XXXXX},   % update once arXiv assigns the ID
  archivePrefix = {arXiv},
  primaryClass = {cs.CL}
}

Please also cite IndicTrans2 (Gala et al., TMLR 2023).

License

MIT (inherited from the IndicTrans2 base model).