bangla-support-qwen3-8b

🇧🇩 BanglaSupport-LLM: Fine-Tuned Bangla Customer Support Model

BanglaSupport-LLM is a domain-adapted, fine-tuned Large Language Model optimized specifically for Bangla E-Commerce Customer Support. Fine-tuned from Qwen2.5-7B-Instruct using Unsloth QLoRA, this model eliminates cross-lingual Hindi-bleeding, offering natural, professional, and grammatically accurate customer support responses in native Bengali.

📌 Model Details

• Developed by: Mahmudur Rahman (mrshibly) & FHJibon (FHJibon)
• Model Type: PEFT / LoRA Adapter for Causal Language Modeling
• Language(s): Bengali (bn), English (en)
• License: MIT
• Finetuned from model: unsloth/Qwen2.5-7B-Instruct-bnb-4bit

🔗 Model Sources & Links

• GitHub Repository: github.com/mrshibly/BanglaSupport-LLM
• Hugging Face Model Hub: huggingface.co/mrshibly/bangla-support-qwen3-8b
• Live Demo (Hugging Face Space): huggingface.co/spaces/mrshibly/bangla-support-llm

---

🎯 Uses

Direct Use

• E-commerce customer service automation in Bangla.
• Answering queries regarding order tracking, shipping, return policies, payment options, and refund eligibility.
• Multi-turn conversational support with system persona framing.

Out-of-Scope Use

• Medical, legal, or high-risk financial advice.
• Generating non-Bengali support text where strict monolingual output is expected.

---

💻 How to Get Started with the Model

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE_MODEL = "unsloth/Qwen2.5-7B-Instruct-bnb-4bit"
ADAPTER_ID = "mrshibly/bangla-support-qwen3-8b"

print("Loading model and tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    device_map="auto",
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(base_model, ADAPTER_ID)
model.eval()

system_prompt = "তুমি একজন সহায়ক বাংলা ই-কমার্স গ্রাহক সেবা সহকারী।"
user_question = "আমার অর্ডারটি ৩ দিন ধরে পেন্ডিং আছে, ডেলিভারি কখন পাব?"

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_question},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=150, temperature=0.7)
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)

print("Response:", response)

---

🏋️ Training Details

Training Data

Trained on a curated dataset of 25,000 normalized Bangla instruction pairs filtered from:

1. md-nishat-008/Bangla-Instruct (ACL 2025 benchmark dataset)
2. CohereForAI/aya_dataset (Bengali subset)

Dataset preprocessing included NFC Unicode normalization, MinHash LSH deduplication, and instruction-intent filtering.

Training Procedure

• Framework: PyTorch 2.11.0 + Unsloth FastLanguageModel + SFTTrainer
• Method: QLoRA 4-bit (NormalFloat4 quantization)
• Precision: bfloat16
• LoRA Parameters: $r = 16$, $\alpha = 32$, Dropout = 0.0
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

---

📊 Evaluation Results

Evaluated against held-out test data using automated metrics & LLM-as-a-Judge benchmarking:

BERTScore evaluated using sagorsarker/bangla-bert-base.

---

🖥️ Hardware & Compute

• Hardware: NVIDIA GeForce RTX 5060 Ti (16GB VRAM)
• Platform: Windows / CUDA 12.0
• Framework Versions: PEFT 0.19.1, Transformers 5.5.0, Unsloth 2026.7.3

---

👨‍💻 Authors & Contact

• Mahmudur Rahman (mrshibly)

  • Email: [email protected]
  • GitHub: @mrshibly
  • Hugging Face: @mrshibly

• FHJibon

  • GitHub: @FHJibon