A LoRA adapter for openbmb/MiniCPM-V-2_6 that turns a document/receipt image into structured line-item JSON ({"menu": [{"nm", "cnt", "price"}], "total": ...}). Built for the Quillwright project (a small-model agent that drafts trade estimates) as the document-extraction skill behind its Document Capture path.

Results — baseline vs. tuned (held-out CORD test split, n=100)

Field-level accuracy of the un-tuned base model vs. this LoRA adapter, scored on the 100-receipt held-out CORD-v2 test split. Deterministic greedy decoding; 0 generation failures. Reproducible via the eval harness in the Quillwright repo (finetune/eval.py).

Metric	Baseline (un-tuned)	Tuned (this LoRA)	Δ
Item F1	0.588	0.681	+0.093
Quantity accuracy	0.715	0.782	+0.067
Price accuracy	0.575	0.726	+0.151
Precision	0.567	0.666	+0.099
Recall	0.647	0.728	+0.081

Every field improved; the largest gain is price accuracy (+0.151) — the field that matters most for an estimate.

Training

Base: openbmb/MiniCPM-V-2_6 (8B vision-language model)
Dataset: naver-clova-ix/cord-v2 (CC BY 4.0, © NAVER CLOVA) — 800-receipt train split; held-out 100-receipt test split for eval
Recipe: OpenBMB's official finetune.py + CPMTrainer, single GPU (L40S), no DeepSpeed, bf16 LoRA (not 4-bit). LoRA on the LLM self-attention projections (q/k/v/o) only; vision tower + resampler frozen (embed_tokens + resampler saved).
Hyperparameters: r=64, α=64, dropout=0.05, lr=1e-5, model_max_length=2048, 3 epochs, effective batch 8 (bs 1 × grad-accum 8), gradient checkpointing.
Final train loss: ~0.31 (from ~0.85).
Attention: SDPA (flash-attn not required).

Inference

python

from peft import PeftModel
from transformers import AutoModel, AutoTokenizer
from PIL import Image

PROMPT = ('Extract the line items from this receipt as JSON with this exact shape: '
          '{"menu": [{"nm": <item name>, "cnt": <quantity>, "price": <price>}], '
          '"total": <grand total>}. Output only the JSON.')

base = AutoModel.from_pretrained("openbmb/MiniCPM-V-2_6", trust_remote_code=True,
                                 attn_implementation="sdpa")
model = PeftModel.from_pretrained(base, "Aarya2004/minicpmv-cord-lora",
                                  trust_remote_code=True).eval().cuda()
tok = AutoTokenizer.from_pretrained("openbmb/MiniCPM-V-2_6", trust_remote_code=True)

img = Image.open("receipt.jpg").convert("RGB")
msgs = [{"role": "user", "content": [img, PROMPT]}]
print(model.chat(image=None, msgs=msgs, tokenizer=tok, sampling=False))

MiniCPM-V-2_6's remote code hard-imports flash_attn at load even with SDPA; if you hit that ImportError, strip flash_attn from transformers.dynamic_module_utils.get_imports (see finetune/flash_patch.py in the Quillwright repo) — flash-attn is not required.

Attribution

Fine-tune recipe adapted from OpenBMB's official MiniCPM-V finetune scripts (Apache-2.0).