This model is a fine-tuned version of Qwen/Qwen3.5-4B. It was trained with LoRA using Unsloth and TRL for English → Kannada translation on the pavan-naik/mt-nllb-en-kn dataset.

Quick start

python

import torch
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="kshitijdesai99/Qwen-3.5-4B-finetuned_mt-nllb-en-kn",
    max_seq_length=512,
    dtype=torch.bfloat16,
    load_in_4bit=False,
)
FastLanguageModel.for_inference(model)

messages = [
    {
        "role": "system",
        "content": "You are a helpful assistant. Translate the given English text to Kannada. Output only the Kannada translation, nothing else.",
    },
    {"role": "user", "content": "Translate to Kannada: Please intercede for us all and especially at the hour of our death, Amen."},
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)

inputs = tokenizer(text=[prompt], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128, do_sample=False)
decoded = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(decoded.strip())

Model Details

Base model: Qwen/Qwen3.5-4B
Task: English → Kannada translation
Method: LoRA (not QLoRA)
Quantization: none
Precision: bfloat16
Hardware: NVIDIA H100 NVL
Dataset: pavan-naik/mt-nllb-en-kn
Splits: 16,702 train / 4,017 validation

Training procedure

This model was trained with supervised fine-tuning (SFT) using LoRA adapters on top of the full-precision Qwen/Qwen3.5-4B base model.

Training hyperparameters

LoRA rank: 16
LoRA alpha: 16
LoRA dropout: 0.0
Train batch size: 128
Eval batch size: 128
Gradient accumulation: 2
Effective batch size: 256
Epochs: 5
Learning rate: 2e-4
Warmup ratio: 0.05
Logging / eval / save steps: 10 / 100 / 100

Training results

Final train loss: 0.4130
Best eval loss: 0.5300 at step 200
Final eval loss: 0.5546
Training runtime: 4660.39s (~1h 17m 40s)
Train samples/sec: 17.919
Train steps/sec: 0.071

Evaluation

Validation set evaluation was run on all 4,017 examples.

Model	chrF2
Zero-shot `Qwen/Qwen3.5-4B`	`45.66`
Zero-shot `Qwen/Qwen3.5-9B`	`51.45`
Fine-tuned `Qwen/Qwen3.5-0.8B`	`48.15`
This LoRA fine-tuned model	`60.20`

Absolute improvement over zero-shot 4B: +14.54 chrF2

Absolute improvement over fine-tuned 0.8B: +12.05 chrF2

Absolute improvement over zero-shot 9B: +8.75 chrF2

Notes

Flash Attention 2 was active during this completed run.
The adapter was trained with standard LoRA, not QLoRA.
The best checkpoint by eval loss was at step 200, while the final saved adapter achieved the best measured translation quality in this project so far.

Framework versions

PEFT 0.19.0
TRL: 0.24.0
Transformers: 5.5.0
Pytorch: 2.10.0
Datasets: 4.3.0
Tokenizers: 0.22.2

Citations

If you use this adapter, please also cite this project repository:

bibtex

@software{desai2026finetuneme,
  author = {Kshitij Desai},
  title = {finetune-me},
  year = {2026},
  url = {https://github.com/kshitijdesai99/finetune-me.git}
}