---
language:
- en
- hi
license: apache-2.0
library_name: transformers
pipeline_tag: translation
tags:
- text-generation-inference
- hinglish
- translation
- code-switching
- llama
- unsloth
- trl
- lora
- text-generation-inference
base_model: unsloth/llama-3-8b-Instruct-bnb-4bit
datasets:
- cmu_hinglish_dog
- suyash2739/Hinglish
---

# Llama 3 8B — English → Hinglish (CMU Hinglish DoG variant)

An earlier variant of my English → Hinglish translation work: Llama 3 8B Instruct fine-tuned with QLoRA on a [cleaned version](https://huggingface.co/datasets/suyash2739/Hinglish) of the CMU Hinglish DoG conversational dataset.

> **Looking for the recommended model?** The newer variant trained on a curated news-domain corpus produces more fluent Hinglish: [English_to_Hinglish_fintuned_lamma_3_8b_instruct](https://huggingface.co/suyash2739/English_to_Hinglish_fintuned_lamma_3_8b_instruct).

## Details

- **Base model:** `unsloth/llama-3-8b-Instruct-bnb-4bit`
- **Method:** QLoRA (4-bit) with Unsloth + HuggingFace TRL
- **Training data:** [suyash2739/Hinglish](https://huggingface.co/datasets/suyash2739/Hinglish) — cleaned from [cmu_hinglish_dog](https://huggingface.co/datasets/cmu_hinglish_dog) (conversational domain)
- **License:** Apache 2.0

## How to use

Same interface as the main model:

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="suyash2739/English_to_Hinglish_cmu_hinglish_dog",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)
```

Prompt format: `Translate the input from English to Hinglish to give the response.` followed by `### Input:` and `### Response:` sections.

## Why two variants?

This model captures conversational, dialogue-style Hinglish (CMU DoG is a document-grounded conversation dataset), while the main model targets news-register Hinglish. Comparing the two illustrates how strongly domain of the parallel corpus shapes code-mixing style in the output.

## Limitations

- Conversational-domain training data; formal text may translate awkwardly.
- Romanized Hinglish only.
- Inherits base-model and corpus biases.