Receipt Form Reader

Status: Live.

1. Your muse

   Rebuild ofFine-tuning Llama models for structured image data extraction (opens in a new tab)Reframe

   Make a tool that pulls structured fields out of photos of receipts and forms, with no fine-tuning. This rebuilds Meta's video "Fine-tuning Llama models for structured image data extraction" (https://dev.meta.ai/llama/resources/videos/llama-fine-tune-structured-image-extraction) on current Muse models. It demonstrates: Customizing Llama to pull structured data out of images. Plan: Same job without fine-tuning. Spark 1.3 reads the image and returns JSON that matches a schema. Upload a receipt or a form and get validated fields back.

2. Show thinking

3. Checks passed

4. 

   The plan: Receipt Form Reader

   Snap a receipt or form and get validated structured fields back.

   1. Upload a photo of one receipt or form
   2. Pick a preset schema or paste a custom JSON schema
   3. Run extraction on the image
   4. Review validated fields and fix flagged values

   See the full plan

   Who it's for

   Freelancers and small teams who need receipt and form data as JSON

   How it works

   1. Upload a photo of one receipt or form
   2. Pick a preset schema or paste a custom JSON schema
   3. Run extraction on the image
   4. Review validated fields and fix flagged values
   5. Copy or download the clean JSON

   Screens

   • Upload and Schema. Upload photo, pick receipt or form preset, or paste custom schema.
   • Results and Validation. Show extracted table and JSON, highlight missing fields, allow copy and download. Footer states built with Muse Muse and not an official Meta product.

   What goes in

   • Image: JPG or PNG photo of one receipt or form. Keep under 4.5 MB or use temporary upload.
   • Text (optional): Optional preset choice or custom JSON schema pasted as text.

   What comes out

   • JSON: Validated fields matching the schema plus flags for low confidence or missing fields.

   Models

   Model	Used for	Why
   Muse Spark 1.3	Read receipt or form photo and return fields that match the schema	Reads images directly and returns schema matched JSON without fine tuning.

   Limits

   • Rebuild fit: reframe. The original video fine tuned Llama for extraction. This tool uses Spark 1.3 with schemas and no training.
   • Works on one clear photo at a time. Blur, glare and handwriting lower accuracy.
   • Results are model readings, not certified OCR. User should check totals before accounting use.
   • No fine tuning, no custom model hosting, and no connection to outside services.

   From your inspo

   Rebuild of Meta video Fine-tuning Llama models for structured image data extraction. Keeps the receipt and form extraction job but does it with prompts and schemas and no fine tuning.

   Takes

   • The job of pulling structured data from receipt and form images into validated fields

   Leaves out

   • Video footage
   • Video script
   • Meta and Llama names and logos

   Cost per run: about $0.006 (est.). About 2500 input tokens at 1.25 per million is 0.0031. About 700 output tokens at 4.25 per million is 0.0030. Total about 0.006.

6. Done. npm run verify exits 0.

   It's live.

   musemuse-receipt-form-reader.vercel.app (opens in a new tab)

   Product pageNew muse

   6 m 8 s · Input 4.2M · Output 52.6k · $1.15 est.