Peter-A-P

legible-4b-cord-v1120-s0-7695d3c79b

README

License: apache-2.0

What it reads and what it writes

Input: the document's pages as images, in page order, then the instruction every reader in the project is sent (legible.prompt, version 1), which describes each field and gives the JSON Schema. The adapter was trained on that instruction and nothing else.

One JSON object a document, every value a string or null, in the compact form it was trained on ({"advertiser":null}, no spaces). A photographed restaurant or shop receipt from Indonesia.

Table with columns: Field, Kind, What it is
Field	Kind	What it is
`menu`	list	One item bought. Each entry has:
`menu[].nm`	string	The item's name.
`menu[].num`	string	The item's code or number, if printed.
`menu[].unitprice`	price	The price of one unit.
`menu[].cnt`	number	How many were bought.
`menu[].discountprice`	price	A discount on this item.
`menu[].price`	price	The line's price.
`menu[].itemsubtotal`	price	The item's subtotal, if printed separately.
`menu[].vatyn`	string	Whether VAT applies to the item, as printed.
`menu[].etc`	string	Anything else printed on the item's line.
`sub_item`	list	A component or option printed under an item, such as a set meal's parts. Each entry has:
`sub_item[].nm`	string	The component's name.
`sub_item[].unitprice`	price	The price of one unit.
`sub_item[].cnt`	number	How many.
`sub_item[].price`	price	The component's price.
`sub_item[].etc`	string	Anything else printed on its line.
`void_menu`	list	An item voided on the receipt. Each entry has:
`void_menu[].nm`	string	The voided item's name.
`void_menu[].price`	price	Its price.
`subtotal_price`	price	The subtotal.
`discount_price`	price	A discount on the whole bill.
`service_price`	price	The service charge.
`othersvc_price`	price	Any other service charge.
`tax_price`	price	The tax.
`subtotal_etc`	string	Anything else printed in the subtotal section.
`total_price`	price	The total to pay.
`total_etc`	string	Anything else printed in the total section.
`cashprice`	price	The cash tendered.
`changeprice`	price	The change given.
`creditcardprice`	price	The amount paid by card.
`emoneyprice`	price	The amount paid by e-money.
`menutype_cnt`	number	The number of distinct items, if printed.
`menuqty_cnt`	number	The total quantity of items, if printed.

How it was trained

Project 06's QLoRA recipe with a vision encoder (legible/train/vision.py): one user turn of page images then the instruction, one assistant turn of the truth's JSON, the loss on the answer's tokens only. Every training document is from the split's train part; a held-out document refuses the set (legible/train/dataset.py).

Table with columns: Setting, Value
Setting	Value
Base	`google/gemma-4-E4B` at revision `411aa17b749aa952df1359d2dcea73917a544d9a`
Chat template	`google/gemma-4-E4B-it` at revision `ee0ef6023621cff504d758262d4e04895a5af4a2`
Quantisation	NF4 4-bit with double quantisation, bf16 compute; the vision and audio towers, the vision projector and the output layer kept at full precision
LoRA	rank 16, alpha 32, dropout 0.05, on the language model's attention and MLP projections and the vision projector, never the towers
Optimisation	learning rate 0.0002, cosine schedule, warmup 0.03, 1 epoch, 1 document a step, 8 steps accumulated
Visual tokens a page	1,120
Sequence limit	12,288 tokens
Seed	0
Training set	`cord` (800 documents)
Training documents and pages	800 documents, 800 pages
Optimiser steps	100, from 800 documents at 8 a step
Loss and timing	in `run.json`, the training record uploaded with the adapter; it was not on the machine that wrote this card

Training data and its licences

CORD v2 receipts (Clova AI), https://huggingface.co/datasets/naver-clova-ix/cord-v2 at commit 7f0115a4b758a71d6473b8d085751692da2fef98. Terms: CC BY 4.0.

"This work is licensed under a [Creative Commons Attribution 4.0 International License][cc-by]."

Quoted from https://github.com/clovaai/cord README, linking to http://creativecommons.org/licenses/by/4.0/; the Hugging Face card has license: cc-by-4.0 (read 2026-10-16). Attribution: Clova AI's CORD (Park et al., "CORD: A Consolidated Receipt Dataset for Post-OCR Parsing", 2019), used under CC BY 4.0.

The adapter is released under apache-2.0, the base model's licence. Publication: CORD is CC BY 4.0, which permits a derived model with attribution.

Results

On CORD's 100 test receipts, read once, level 0, temperature 0, unconstrained. Field accuracy is VRDU's matchers, one grader for every reader; intervals are 95% bootstrap over documents. The adapter is costed at an RTX 3090 at US$0.50 an hour on RunPod (2026-10-18) over the measured wall time of its read; the frontier models at the gateway ledger's cost at the 2026-10-16 price list, batch rate.

Table with columns: Reader, Fields correct (95% CI), Every field right, Invented fields, Answers that did not parse, Cost per 1,000 pages, Documents
Reader	Fields correct (95% CI)	Every field right	Invented fields	Answers that did not parse	Cost per 1,000 pages	Documents
B, Gemma 4 E4B fine-tuned (this adapter)	77.6% (71.5 to 83.3)	30.0% (21.0 to 39.0)	0.9% (0.4 to 1.5)	0.0% (0.0 to 0.0)	US$0.16 (0.16 to 0.16)	100
C, Gemma 4 E4B untuned on the image	60.0% (54.6 to 65.2)	3.0% (0.0 to 7.0)	9.9% (8.3 to 11.8)	2.0% (0.0 to 5.0)	US$0.19 (0.19 to 0.19)	100
A, Claude Sonnet 5.5 on the image	84.0% (77.8 to 89.2)	47.0% (37.0 to 57.0)	3.4% (2.3 to 4.6)	0.0% (0.0 to 0.0)	US$5.27 (5.06 to 5.50)	100
A, Claude Haiku 5.5 on the image	76.7% (70.1 to 82.6)	28.0% (19.0 to 37.0)	5.2% (4.0 to 6.6)	2.0% (0.0 to 5.0)	US$0.38 (0.36 to 0.40)	100
A, GPT-6.1 Sol on the image	88.1% (84.2 to 91.6)	51.0% (41.0 to 61.0)	3.4% (2.4 to 4.6)	0.0% (0.0 to 0.0)	US$4.59 (4.35 to 4.85)	100
A, Gemini 3.8 Flash on the image	86.1% (79.7 to 91.6)	57.0% (47.0 to 66.0)	1.2% (0.6 to 2.0)	5.0% (1.0 to 10.0)	US$4.42 (3.77 to 5.12)	100

The best frontier model on these documents is GPT-6.1 Sol, at 88.1% (84.2 to 91.6) of fields for US $4.59 (4.35 t o 4.85) a t h o u s an d p a g es . T hi s a d a pt e r' s f i e l d a cc u r a cy d i f f er s f r o mi t b y - 10.5 p o in t s, a t U S$ 0.16 (0.16 to 0.16) a thousand pages, about 1/28 of the cost. Against its untuned base it reads +17.6 points.

Training variance: the same recipe at seeds 1 and 2, each on its own split's test set: seed 1, 75.3% (68.4 to 81.7); seed 2, 76.2% (70.2 to 81.9). The three seeds' point estimates span 2.3 points (their test sets differ where the split files do).

The 03 release gate

Against Claude Haiku 5.5: no decision: every field under-powered at 3 points (fields pass, block, under-powered: 0, 0, 30).
Against Gemini 3.8 Flash: no decision: every field under-powered at 3 points (fields pass, block, under-powered: 0, 0, 30).

Failure modes

A form design it has not seen. Not measured on this corpus: CORD's receipts have no template labels. On FARA and DeepForm, where it was measured, the fine-tunes read worse on templates they did not see and the frontier models do not (the README's template gap tables); expect worse on a form unlike its training set.
A bad scan. Not measured on receipts, which are photographs rather than scans; how it fails under blur or glare is unknown. The README's NIST and VRDU curves show what the degradation steps do to the other fine-tunes.
Confidently wrong. Its own token probability is not a reason to skip review. Of its dev fields written at a token probability of 0.999 or more (167), 77.8% (61.8 to 91.2) are right; the commonest error there is the value of another field on the page (24). At 99% accuracy it sends 100.0% (100.0 to 100.0) of its test fields to a person, since no threshold reached 99% on dev. The probability measures how sure the model is of its habit, not whether the page agrees (docs/review.md).
Training variance. The same recipe at three seeds reads from 75.3% to 77.6% of fields on test (above); a retrained copy is not guaranteed to read like this one.
Not non-inferior to the frontier. Project 03's gate cannot decide on this corpus at a 3-point margin per field (above). Do not take it as a drop-in for a frontier model's read.
Malformed answers. 0.0% (0.0 to 0.0) of its test answers did not parse as JSON. It is read freely: holding the fine-tunes to the schema's grammar was tried on dev and rejected, because it pushed them off their trained spacing and the DeepForm E2B then answered null for every field (docs/rejected.md). A caller must count a parse failure as a document to type by hand.
Invented fields. 0.9% (0.4 to 1.5) of scored fields carry a value where the form's box is blank.
One instruction. Trained on one prompt and one schema; another wording, another field list or another language is untested.

Intended use

Reading CORD receipts of the kind it was trained on into JSON for a person to check, at a fraction of a frontier model's cost per page; and reproducing the measurements in this card. Plan on a person checking every field: no confidence measured in the project separates right fields from wrong ones well enough to skip review at 99% (the README's honest limitation).

Out of scope

Any decision about a person made on its output without a person checking it.
Handwriting, other form families, long multi-page documents, or a form whose type is unknown.
Unattended use on form designs it was not trained on, where it is worst.
Pages that are not public: it was measured on public records and datasets only.

Loading it

python

from peft import PeftModelfrom transformers import AutoModelForImageTextToText, AutoProcessor base = AutoModelForImageTextToText.from_pretrained("google/gemma-4-E4B", revision="411aa17b749aa952df1359d2dcea73917a544d9a")model = PeftModel.from_pretrained(base, "<this repository>")processor = AutoProcessor.from_pretrained("<this repository>")processor.image_processor.max_soft_tokens = 1120

Read it at temperature 0, with the prompt legible.prompt.build_prompt writes for its schema; the project served it merged into the bf16 base under vLLM (legible serve merge).

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Peter-A-P

Model Tree

Base

google/gemma-4-E4B

Adapter

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality