LFM2.5-VL-450M-PharmaOCR-GGUF

LFM2.5-VL-450M-PharmaOCR (GGUF)

GGUF build of a LoRA fine-tune of LiquidAI/LFM2.5-VL-450M-Extract. Reads photographed Indian pharmacy and hospital invoices into a markdown line-item table (name, quantity, MRP/price).

Vision-language model. Runs on llama.cpp through llama-mtmd-cli. A text quant and the mmproj vision file are both required.

Files

Q4_K_M plus the vision projector is 419 MB.

Usage

llama-mtmd-cli \
  -m lfm2vl_v2-Q4_K_M.gguf \
  --mmproj mmproj-lfm2vl_v2-f16.gguf \
  --image invoice.jpg \
  -p "Transcribe this pharmacy/medical invoice as a markdown document, including every line item in the table with its full name, quantity, and MRP/price."

Greedy decoding. Requires a llama.cpp build with Lfm2VlForConditionalGeneration support (b10050 or later).

Training

LoRA on the language tower and the vision-to-language projector. Vision encoder frozen. Trained on synthetic Indian pharmacy and hospital invoices with per-invoice variation in font, ink, header wording, and density. The target is a single layout-agnostic item list shared across all templates. No real patient data was used.

Evaluation

Composite score = 0.55 name + 0.30 strength + 0.075 quantity + 0.075 price, fuzzy-matched against gold. Small internal set of held-out real invoices (n=3).

Post-training quantization is lossless on this set. Q4_K_M is the recommended build.

Intended use and limits

Structured extraction from Indian pharmacy and hospital invoice images. The model was trained on that domain only; behavior on other document types is undefined. A generative reader can drop or misread line items, so its output is not a substitute for the source document in any safety-critical path.

License

LFM Open License v1.0. See LICENSE. Commercial use is limited by an annual-revenue threshold set by the licensor; review the terms before commercial deployment. This model is a derivative of LiquidAI/LFM2.5-VL-450M-Extract and is bound by the same license.