README.md

---
license: apache-2.0
task_categories:
- image-to-text
- visual-question-answering
tags:
- vlm
- unsloth
- qwen-vl
- document-ai
- trade-finance
- bangladesh-bank
size_categories:
- n<1K
---

# 🏦 Bangladesh Bank Trade Finance IDP — Multi-User Collaborative Fine-Tuning Dataset

This dataset contains human-reviewed, verified, and corrected document extractions for the **8 official Bangladesh Bank regulatory trade-finance document types**.

It is **completely self-contained** and structured for immediate Vision-Language Model (VLM) fine-tuning anytime from any environment (Colab, Kaggle, GPU cluster, or local), with built-in multi-annotator merge support and incremental delta uploads.

---

## 📊 Dataset Structure & Splits

| Split / Subset | Description | Examples |
|---|---|---|
| **`train`** | Ready-to-train multi-task VLM conversations (`messages` format) | 320 |
| **`validation`** | Stratified held-out evaluation conversations for monitoring eval loss | 80 |
| **`documents`** | Full document-level feedback records with model outputs, human corrections, confidence scores & DPO pairs (subset: `documents`) | 200 |

### 📑 Document Types Represented
| Document Type | Document Count |
|---|---|
| `air_waybill` | 25 |
| `bill_of_entry` | 25 |
| `commercial_lca` | 25 |
| `exp_form` | 25 |
| `final_invoice` | 25 |
| `imp_form` | 25 |
| `industrial_lca` | 25 |
| `ocean_bl` | 25 |

**Total DPO Preference Pairs Available:** 10 (for Direct Preference Optimization)

- **Visual Diversity:** 8 Header styles, 5 Table formats, 4 Procedural Stamp configurations, and 12 distinct color palettes.
- **Realistic Handwriting:** Natural handwriting pen fills and endorsements in varied inks (blue ballpoint, dark ink, navy, royal blue) paired with 100% strictly validated schema JSON.
- **Header Integrity:** `MessageIdentifier` is strictly excluded from document image layouts and is normalized as empty `""` in ground truth.

---

## 🚀 Quickstart: Train in 10 Lines with Unsloth / TRL

You can fine-tune Qwen3-VL / Qwen2.5-VL directly on this dataset without ANY manual data wrangling:

```python
from datasets import load_dataset
from unsloth import FastVisionModel
from trl import SFTTrainer, SFTConfig

# 1. Load dataset directly from Hugging Face
ds = load_dataset("bisalsaha/bb-trade-idp-feedback")
train_data = ds["train"]
eval_data = ds.get("validation")

# 2. Load model & attach vision LoRA adapters
model, tokenizer = FastVisionModel.from_pretrained(
    "unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit",
    load_in_4bit=True
)
model = FastVisionModel.get_peft_model(
    model,
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"]
)

# 3. SFT Trainer
trainer = SFTTrainer(
    model=model,
    train_dataset=train_data,
    eval_dataset=eval_data,
    dataset_text_field="messages",
    max_seq_length=2048,
    args=SFTConfig(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        learning_rate=1e-4,
        output_dir="./qwen3vl_trade_idp"
    )
)
trainer.train()
```