Qwen 3 0.6B English <-> Hinglish (Preview)
This is just a English to Hinglish (code-switched English-Hindi) and back transaltion model. Fine-tuned from Qwen 3 0.6B (with Unsloth), this model is designed for translation tasks on any device!
Issues
Low BLEU Scores: Because I trained this for ONLY a epoch, BLEU scores are lower than in the final version. We're going to train a new version for 5 epochs!
Notes
This is still a experimental model and shouldn't be used for tasks where accurate translations matter!
Evaluation Results
| Direction | BLEU (250 sentences from test split) |
|---|---|
| English -> Hinglish | 4.94 |
| Hinglish -> English | 17.89 |
Usage
Code is by Gemini 3 Flash (then some little modifications by myself):
English to Hinglish
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
# 1. Load from your Hugging Face Repo
model_id = "MihaiPopa-1/Qwen3-0.6B-English-Hinglish-Preview"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float32, # Standard for CPU
device_map="cpu" # Forces CPU usage
)
# 2. Translate (replace ron_Latn with your language here)
prompt = "<|im_start|>user\nTranslate English to Hinglish: Hello, how are you doing?<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cpu")
with torch.no_grad():
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Hinglish to English
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
# 1. Load from your Hugging Face Repo
model_id = "MihaiPopa-1/Qwen3-0.6B-English-Hinglish-Preview"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float32, # Standard for CPU
device_map="cpu" # Forces CPU usage
)
# 2. Translate (replace ron_Latn with your language here)
prompt = "<|im_start|>user\nTranslate Hinglish to English: Hello, tum kaise ho?<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cpu")
with torch.no_grad():
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Data Used
I used CMU's Hinglish Dog for this model (yes, really)!
Uploaded finetuned model
- Developed by: MihaiPopa-1
- License: apache-2.0
- Finetuned from model : unsloth/qwen3-0.6b-unsloth-bnb-4bit
This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.