xprilion/granite-4.1-3b-hindi-lora

Granite 4.1 3B — Hindi Instruct (LoRA)

A QLoRA fine-tune of ibm-granite/granite-4.1-3b on Hindi instruction data. The model responds to prompts in natural Hindi while retaining its original English and code capabilities.

Blog: xprilion.com/teching-ibm-granite-hindi-on-a-laptop-gpu

🚀 Capabilities

• Conversational Hindi with natural phrasing
• Hindi question-answering (factual + creative)
• Code explanation in Hindi
• Poetry and creative writing in Hindi

📊 Evaluation

Benchmarked on a held-out Hindi instruction set:

Metric	Base Granite 4.1	Fine-tuned
Perplexity	7.30	1.85
Training Loss	1.28	0.53

Trained for 400 steps on an RTX 3070 Laptop GPU (8GB VRAM) using Unsloth QLoRA (r=8, 4-bit). Dataset: FreedomIntelligence/evol-instruct-hindi.

🛠️ Usage

This is a merged 16-bit model — load directly with Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "xprilion/granite-4.1-3b-hindi-lora",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("xprilion/granite-4.1-3b-hindi-lora")

prompt = "भारत की राजधानी क्या है?"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

For 4-bit inference to save VRAM:

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

model = AutoModelForCausalLM.from_pretrained(
    "xprilion/granite-4.1-3b-hindi-lora",
    device_map="auto"
)

📝 Training Details

• Base model: ibm-granite/granite-4.1-3b
• Dataset: FreedomIntelligence/evol-instruct-hindi (59K samples)
• Method: QLoRA (r=8, alpha=16, 4-bit NF4)
• Context length: 512 tokens
• Effective batch size: 8
• Optimizer: AdamW 8-bit, cosine schedule, lr=2e-4
• Framework: Unsloth + SFTTrainer (TRL)

📄 License

Apache 2.0 — same as the base Granite 4.1 model.