RLM-hinglish-translator-bidir-2 is a fully bidirectional Hinglish ↔ English translation model — the first in this series to support translation in both directions. Built on the same 4B instruction-tuned base as v2 and fine-tuned using PEFT (LoRA) with TRL's SFTTrainer, it lets you seamlessly convert between informal Hinglish and English without needing two separate models.

GGUF VERSION

What's New

Bidirectional translation — Translate from Hinglish → English and English → Hinglish in a single model.
Direction-aware prompting — The model uses explicit instruction prompts to distinguish translation direction at inference time.
Same strong base — Retains the 4B instruction-tuned backbone from v2 for high fluency and contextual understanding.
Expanded training — Fine-tuned on both directions of the findnitai/english-to-hinglish dataset for balanced performance.

Usage

Hinglish → English

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "rudrashah/RLM-hinglish-translator-bidir-2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

messages = [
    {
        "role": "user",
        "content": "Hinglish:\naapka name kya hai?"
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    use_cache=True,
    temperature=1.0,
    top_p=0.95,
    top_k=64,
)
print(tokenizer.batch_decode(outputs))

English → Hinglish

python

messages = [
    {
        "role": "user",
        "content": "English:\nWhat are you doing tomorrow?"
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    use_cache=True,
    temperature=1.0,
    top_p=0.95,
    top_k=64,
)
print(tokenizer.batch_decode(outputs))

Note: Always include the Language of Input for Best Result.

Examples

Hinglish → English

Hinglish Input	English Output
`aapka name kya hai?`	`What is your name?`
`kal mujhe office jaana hai`	`I have to go to the office tomorrow.`
`Tumne kal ki match dekhi thi? kon man of the match bana tha?`	`Did you watch yesterday's match? Who was man of the match?`

English → Hinglish

English Input	Hinglish Output
`What is your name?`	`aapka naam kya hai?`
`I have to go to the office tomorrow.`	`kal mujhe office jaana hai.`
`Did you watch yesterday's match?`	`Tumne kal ki match dekhi thi?`

Limitations

Best suited for short to medium-length informal sentences, similar to everyday chat messages.
May struggle with heavy regional slang, very long sentences, or dialect-specific expressions.
English → Hinglish outputs reflect the Hinglish style present in the training data and may not match all regional variations.

Model History

Model	Direction	Base Model	Notes
RLM-hinglish-translator	Hinglish → English	2B	Original release
RLM-hinglish-translator-2	Hinglish → English	4B	Chat-format fine-tuning, LoRA rank 64
RLM-hinglish-translator-bidir-2 (this)	Hinglish ↔ English	4B	Bidirectional support

Thanks to that one girl who motivated me. Thank you P.