Bangladesh Bank Trade Finance IDP Model
Fine-tuned Qwen3-VL for extracting the 8 official Bangladesh Bank trade/FX document types (commercial/industrial LCA, Bill of Entry, EXP/IMP forms, final invoice, air waybill, ocean B/L) into regulatory XML.
Overall mean field accuracy: 0.0%
| Document Type | N | Field Acc | Exact Match |
|---|---|---|---|
| air_waybill | 1 | 0.0% | 0.0% |
| bill_of_entry | 1 | 0.0% | 0.0% |
| commercial_lca | 1 | 0.0% | 0.0% |
| exp_form | 1 | 0.0% | 0.0% |
| final_invoice | 1 | 0.0% | 0.0% |
| imp_form | 1 | 0.0% | 0.0% |
| industrial_lca | 1 | 0.0% | 0.0% |
| ocean_bl | 1 | 0.0% | 0.0% |
Load the merged model (no PEFT needed)
from transformers import AutoModelForVision2Seq, AutoProcessor
model = AutoModelForVision2Seq.from_pretrained("jihadv4/bb-trade-idp-model", torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained("jihadv4/bb-trade-idp-model")
Load the adapter (with Unsloth base)
from unsloth import FastVisionModel
model, tokenizer = FastVisionModel.from_pretrained("jihadv4/bb-trade-idp-model", load_in_4bit=True)
FastVisionModel.for_inference(model)
- Downloads last month
- 11
Model tree for jihadv4/bb-trade-idp-model
Base model
Qwen/Qwen3-VL-8B-Instruct