lubzo/marathi-alpaca-cleaned-translated

Marathi Alpaca Cleaned Translated

A Marathi translation of the 51,760-row Alpaca-Cleaned instruction-tuning dataset — Unsloth's hosted fork of yahma/alpaca-cleaned, which fixes hallucinations, empty outputs, and formatting errors found in the original Stanford Alpaca-52k dataset.

Translated using Meta's facebook/nllb-200-distilled-600M model. Built to reproduce and evaluate the Marathi instruction-tuning experiment from Khade et al., CHiPSAL 2025. The original paper translated the uncleaned Alpaca-52k using Google Translate; this version starts from the cleaned source and uses an open translation model.

---

Dataset Description

Each row contains an (instruction, input, output) triple from unsloth/alpaca-cleaned, with all three fields translated into Marathi. The input field is often empty, matching the structure of the source dataset.

• Source dataset: unsloth/alpaca-cleaned (51,760 rows, English)
• Translation model: facebook/nllb-200-distilled-600M
• Source language code: eng_Latn
• Target language code: mar_Deva (Marathi)
• Rows translated: 51,760 / 51,760
• Companion GitHub repository: lubzo/gemma-marathi-lora-reproduction

---

Dataset Creation & Translation Details

Translation Scripts

Translation was executed primarily with CTranslate2, with a reference PyTorch implementation also included in the repository:

1. Primary Script (src/translate_ct2.py):
   • Uses the CTranslate2 engine (float16 / int8 compute types) with greedy decoding (beam_size=1) and target prefix token mar_Deva.
2. Reference Script (src/translate.py):
   • Standard Hugging Face transformers implementation using AutoModelForSeq2SeqLM (torch.float16) with greedy search.

Token Limits

Translation was performed with the following generation limits:

• instruction: max_decoding_length = 256 tokens
• input: max_decoding_length = 256 tokens
• output: max_decoding_length = 512 tokens

No manual review or post-editing was performed on the translated text.

---

Note on Reproduction Fidelity

The original paper (Khade et al., CHiPSAL 2025) translated uncleaned tatsu-lab/alpaca using Google Translate.

Using unsloth/alpaca-cleaned and nllb-200-distilled-600M is a deliberate deviation:

• Starts from cleaned data with fewer upstream errors.
• Uses an open-weights translation model instead of a commercial API.

---

Dataset Structure

{
  "instruction": "खालील शब्दांचे विरुद्धार्थी शब्द सांगा.",
  "input": "आनंद",
  "output": "आनंद या शब्दाचा विरुद्धार्थी शब्द 'दुःख' असा आहे."
}

Field	Type	Description
instruction	string	The instruction or task prompt, translated to Marathi (mar_Deva).
input	string	Optional additional context, translated to Marathi (often empty "").
output	string	The expected response, translated to Marathi (mar_Deva).

---

How to Use

from datasets import load_dataset

# Load the translated dataset
ds = load_dataset("lubzo/marathi-alpaca-cleaned-translated")

# Inspect a training example
print(ds["train"][0])

Example Prompt Formatting

def format_prompt(example):
    instruction = example["instruction"]
    user_input = example.get("input", "").strip()
    response = example["output"]
    
    if user_input:
        prompt = f"### Instruction:\n{instruction}\n\n### Input:\n{user_input}\n\n### Response:\n{response}"
    else:
        prompt = f"### Instruction:\n{instruction}\n\n### Response:\n{response}"
        
    return {"text": prompt}

formatted_ds = ds["train"].map(format_prompt)

---

Limitations

1. Machine-Translated: This is raw machine translation without human review. Expect translation artifacts, occasional awkward phrasing, or literal translations of idioms.
2. Translation Model Size: Translated with nllb-200-distilled-600M for practical compute constraints; larger checkpoints or dedicated Indic translation models may produce higher-quality output.
3. No Filtering Applied: No quality, length-ratio, or semantic filters were applied to the translations.
4. Upstream Dataset: The data derives from Alpaca and naturally carries any foundational biases or limitations present in the original dataset.

---

Licensing & Terms of Use

This dataset combines two upstream sources:

1. Source Text (unsloth/alpaca-cleaned): Derived from Stanford Alpaca-52k, generated via OpenAI's API. Governed by OpenAI's Terms of Use (non-commercial/research use). While the repository tag lists cc-by-4.0, the dataset card specifies CC-BY-NC 4.0.
2. Translation Model (facebook/nllb-200-distilled-600M): Released by Meta AI under CC-BY-NC 4.0.

Takeaway: This dataset is released under CC-BY-NC 4.0 for non-commercial research use.

---

Citation

@inproceedings{khade2025challenges,
  title={Challenges in Adapting Multilingual LLMs to Low-Resource Languages using LoRA PEFT Tuning},
  author={Khade, Omkar and Jagdale, Shruti and Phaltankar, Abhishek and Takalikar, Gauri and Joshi, Raviraj},
  booktitle={Proceedings of the First Workshop on Challenges in Processing South Asian Languages (CHiPSAL 2025)},
  pages={217--222},
  year={2025}
}

@misc{alpaca-cleaned,
  author = {Gene Ruebsamen},
  title = {Alpaca-Cleaned: A curated, cleaned version of the Alpaca dataset},
  year = {2023},
  howpublished = {\url{https://github.com/gururise/AlpacaDataCleaned}}
}