README.md

---
base_model: unsloth/Qwen2.5-7B-Instruct-bnb-4bit
library_name: peft
pipeline_tag: text-generation
language:
- bn
- en
tags:
- base_model:adapter:unsloth/Qwen2.5-7B-Instruct-bnb-4bit
- lora
- sft
- transformers
- trl
- bangla
- customer-support
license: mit
---

# 🇧🇩 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](https://github.com/mrshibly)) & Ferdous Hasan ([FHJibon](https://github.com/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/FHJibon/BanglaLLM](https://github.com/FHJibon/BanglaLLM)
- **Hugging Face Model Hub:** [huggingface.co/FHJibon/Bangla-LLM](https://huggingface.co/FHJibon/Bangla-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

```python
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:

| Model Variant | BLEU-4 | ROUGE-L | BERTScore (F1) | LLM-Judge (Fluency) | LLM-Judge (Accuracy) |
|---|:---:|:---:|:---:|:---:|:---:|
| Base Qwen2.5-7B-Instruct | 0.1820 | 0.3840 | 0.7620 | 3.4 / 5.0 | 3.1 / 5.0 |
| **Fine-Tuned BanglaSupport-LLM** | **0.4280** | **0.6910** | **0.9140** | **4.8 / 5.0** | **4.7 / 5.0** |

*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)**
  - **GitHub:** [@mrshibly](https://github.com/mrshibly)

- **FHJibon**
  - **GitHub:** [@FHJibon](https://github.com/FHJibon)