HimangY/BaatCheet-Corpus

BaatCheet: A Multilingual Corpus for Dialogue Translation in Indian Languages

BaatCheet (बातचीत) — named after the Hindi term for conversation or chitchat — is a multilingual dialogue translation corpus containing approximately 49,000 dialogues across 5 translation directions in Indian languages.

📖 Abstract

Existing translation models are typically trained on sentence-level and formal text, limiting their ability to capture everyday conversational dialogue phenomena such as informality, speaker interaction, and discourse coherence. Most existing Indic translation resources and evaluation benchmarks focus on sentence-level or formal text, making it difficult to assess translation quality of the dialogue phenomena.

In this work, we introduce BaatCheet, a multilingual dialogue corpus named after the Hindi term for conversation or chitchat, containing approximately 49,000 dialogues for dialogue translation across five translation directions. We fine-tune five open-source LLMs across seven training data configurations and find that fine-tuning yields substantial gains over zero- and few-shot baselines. To comprehensively evaluate dialogue translation quality, we employ multiple evaluation strategies, including automatic metrics, LLM-as-judge, and human assessments using an SQM-guided Direct Assessment (DA) Protocol.

---

🌐 Language Directions

BaatCheet covers 5 translation directions across English and Indic languages (Hindi, Telugu, Tamil):

1. English (eng) → Hindi (hin)
2. English (eng) → Telugu (tel)
3. English (eng) → Tamil (tam)
4. Hindi (hin) → Telugu (tel)
5. Hindi (hin) → Tamil (tam)

---

📊 Overall Corpus Statistics

Below are the overall statistics of the BaatCheet corpus across its 4 splits:

Subset	#Dialogues	#Utterances	#Sent (Src)	#Sent (Tgt)	Avg Dlg Len	Avg Utt Len (Src)	Avg Utt Len (Tgt)	Avg Sent Len (Src)	Avg Sent Len (Tgt)
BaatCheet_Syn_X2IL	42,148	301,278	533,102	533,494	8.8	14.8	11.8	8.7	6.9
BaatCheet_Syn_IL2X	6,096	73,666	135,894	136,616	12.3	14.6	11.8	8.4	6.9
BaatCheet_Human_Train	510	8,405	16,455	16,077	17.3	17.6	14.1	9.1	7.5
BaatCheet_Test	658	8,904	16,813	16,518	13.7	16.9	14.1	8.9	7.5
TOTAL	49,412	392,253	702,264	702,705	13.0	16.0	13.0	8.8	7.2

---

📁 Dataset Organization & Configs

The dataset is organized into 5 configs (one per language pair) and 2 HuggingFace splits (train and gold):

• train split: Combines synthetic training subsets (BaatCheet_Syn_I and BaatCheet_Syn_II).
• gold split: Combines gold human-annotated training and test sets (BaatCheet_Human_Train and BaatCheet_Test).

Config	train Split	gold Split
eng_hin	BaatCheet_Syn_I/eng_hin + BaatCheet_Syn_II/eng_hin	BaatCheet_Human_Train/eng_hin + BaatCheet_Test/eng_hin
eng_tel	BaatCheet_Syn_I/eng_tel + BaatCheet_Syn_II/eng_tel	BaatCheet_Human_Train/eng_tel + BaatCheet_Test/eng_tel
eng_tam	BaatCheet_Syn_I/eng_tam + BaatCheet_Syn_II/eng_tam	BaatCheet_Human_Train/eng_tam + BaatCheet_Test/eng_tam
hin_tel	BaatCheet_Syn_I/hin_tel + BaatCheet_Syn_II/hin_tel	BaatCheet_Human_Train/hin_tel + BaatCheet_Test/hin_tel
hin_tam	BaatCheet_Syn_I/hin_tam + BaatCheet_Syn_II/hin_tam	BaatCheet_Human_Train/hin_tam + BaatCheet_Test/hin_tam

---

📁 Directory Structure

BaatCheet_Corpus/
├── BaatCheet_Syn_I/                ← Syn_X2IL synthetic dialogues (English source)
│   ├── eng_hin_tel_tam.tsv         ← Multi-target (eng | hin | tel | tam)
│   ├── hin_tel_tam.tsv             ← Multi-target (hin | tel | tam)
│   ├── eng_hin/ eng_hin.tsv
│   ├── eng_tel/ eng_tel.tsv
│   ├── eng_tam/ eng_tam.tsv
│   ├── hin_tel/ hin_tel.tsv
│   └── hin_tam/ hin_tam.tsv
│
├── BaatCheet_Syn_II/               ← Syn_IL2X synthetic dialogues (Indic source)
│   ├── eng_hin/ eng_hin.tsv
│   ├── eng_tel/ eng_tel.tsv
│   ├── eng_tam/ eng_tam.tsv
│   ├── hin_tel/ hin_tel.tsv
│   └── hin_tam/ hin_tam.tsv
│
└── BaatCheet_Gold/
    ├── BaatCheet_Human_Train/      ← Gold human-annotated train dialogues
    │   ├── eng_hin_tel_tam.tsv     ← Multi-target (eng | hin | tel | tam)
    │   ├── hin_tel_tam.tsv         ← Multi-target (hin | tel | tam)
    │   ├── eng_hin/ eng_hin.tsv
    │   ├── eng_tel/ eng_tel.tsv
    │   ├── eng_tam/ eng_tam.tsv
    │   ├── hin_tel/ hin_tel.tsv
    │   └── hin_tam/ hin_tam.tsv
    │
    └── BaatCheet_Test/             ← Gold benchmark test set
        ├── eng_hin_tel_tam.tsv     ← Multi-target (eng | hin | tel | tam)
        ├── hin_tel_tam.tsv         ← Multi-target (hin | tel | tam)
        ├── eng_hin/ eng_hin.tsv
        ├── eng_tel/ eng_tel.tsv
        ├── eng_tam/ eng_tam.tsv
        ├── hin_tel/ hin_tel.tsv
        └── hin_tam/ hin_tam.tsv

---

💻 Usage & Loading Code

Using HuggingFace datasets

from datasets import load_dataset

# Load English -> Hindi train split (Syn_I + Syn_II)
dataset_train = load_dataset("HimangY/BaatCheet_Corpus", name="eng_hin", split="train")
print(dataset_train[0])

# Load Hindi -> Telugu gold split (Human_Train + Test)
dataset_gold = load_dataset("HimangY/BaatCheet_Corpus", name="hin_tel", split="gold")
print(dataset_gold[0])

Using Pandas

import pandas as pd

# Load multi-target TSV for English source
df_multi = pd.read_csv("BaatCheet_Syn_I/eng_hin_tel_tam.tsv", sep="\t", keep_default_na=False)
print(df_multi.head())

# Load individual pair TSV
df_pair = pd.read_csv("BaatCheet_Gold/BaatCheet_Human_Train/eng_hin/eng_hin.tsv", sep="\t", keep_default_na=False)
print(df_pair.head())

---

📜 Citation & License

@inproceedings{baatcheet2026,
  title={BaatCheet: A Multilingual Corpus for Dialogue Translation in Indian Languages},
  author={...},
  booktitle={...},
  year={2026}
}

This corpus is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.