fixensydev/bb-trade-idp-feedback

🏦 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)	264
validation	Stratified held-out evaluation conversations for monitoring eval loss	56
documents	Full document-level feedback records with model outputs, human corrections, confidence scores & DPO pairs (subset: documents)	160

📑 Document Types Represented

Document Type	Document Count
air_waybill	18
bill_of_entry	22
commercial_lca	20
exp_form	19
final_invoice	23
imp_form	19
industrial_lca	20
ocean_bl	19

Total DPO Preference Pairs Available: 9 (for Direct Preference Optimization)

---

🚀 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:

from datasets import load_dataset
from unsloth import FastVisionModel
from unsloth.trainer import UnslothVisionDataCollator
from trl import SFTTrainer, SFTConfig

# 1. Load the ready-to-train dataset directly from Hugging Face
ds = load_dataset("jihadv4/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-Max-Instruct",
    load_in_4bit=True,
    max_seq_length=2048,
)
model = FastVisionModel.get_peft_model(
    model,
    finetune_vision_layers=True,
    finetune_language_layers=True,
    finetune_attention_modules=True,
    finetune_mlp_modules=True,
    r=16,
    lora_alpha=16,
    lora_dropout=0.05,
)
FastVisionModel.for_training(model)

# 3. Train with SFTTrainer
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    data_collator=UnslothVisionDataCollator(model, tokenizer),
    train_dataset=train_data,
    eval_dataset=eval_data,
    args=SFTConfig(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        learning_rate=1e-4,
        num_train_epochs=3,
        optim="adamw_8bit",
        output_dir="./lora_checkpoints",
    ),
)
trainer.train()

---

📦 Standalone Artifacts in this Repository

• images/: Content-addressed directory containing every verified document page image (LFS deduplicated).
• images_backup.tar.gz: Standalone compressed archive of all page images for raw PyTorch / non-parquet workflows.
• feedback.jsonl: Exact raw JSONL feedback log with per-record batch_id and synced_at watermarks.
• schemas_bundle.json: Official Bangladesh Bank JSON schemas, field definitions, and XML templates.
• trade_ontology.json: Concept synonym mappings and bilingual hints.
• training_config.json: Recommended hyperparameters, target modules, and runtime compatibility settings.