Peter-A-P

legible-4b-deepform-unseen-s0-v1120-s0-27c2e83f69

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 political advertising order, invoice or contract from a US television station's FCC public file.

Table with columns: Field, Kind, What it is
Field	Kind	What it is
`advertiser`	string	The advertiser buying the airtime.
`agency`	string	The agency that placed the order, if any.
`contract_num`	code	The contract or order number.
`flight_from`	date	The first date of the flight (the run of the ads).
`flight_to`	date	The last date of the flight.
`gross_amount`	price	The gross total of the order.
`product`	string	The product, candidate or issue advertised.
`property`	string	The station's call letters.
`tv_address`	address	The station's postal address.
`line_item`	list	One scheduled row of the order. Each entry has:
`line_item[].channel`	string	The channel the row airs on.
`line_item[].program_desc`	string	The programme or daypart.
`line_item[].program_start_date`	date	The row's first air date.
`line_item[].program_end_date`	date	The row's last air date.
`line_item[].sub_amount`	price	The row's amount.

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	`deepform-unseen-s0` (200 documents)
Training documents and pages	200 documents, 476 pages
Optimiser steps	25, from 200 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

VRDU Ad-buy Forms (DeepForm), https://github.com/google-research-datasets/vrdu (ad-buy-form). Terms: CC BY 4.0 for the benchmark's annotations; the pages are US public records, political advertising orders from television stations' FCC public files.

"We are excited to announce the public release of the VRDU dataset and evaluation code under a Creative Commons license."

Quoted from Google's release post, "Advances in document understanding" (2023-08-09), https://research.google/blog/advances-in-document-understanding/, where "Creative Commons license" links to https://creativecommons.org/licenses/by/4.0/. That hyperlink is the only statement of the variant; the repository has no licence file (read 2026-10-16). Attribution: Google Research's VRDU benchmark (Wang et al., "VRDU: A Benchmark for Visually-rich Document Understanding", KDD 2023), annotations used under CC BY 4.0.

The adapter is released under apache-2.0, the base model's licence. Publication: VRDU's annotations are CC BY 4.0 and the pages are public records from the FCC's public files; neither forbids a derived model.

Results

On the 159 test documents of VRDU's split file DeepForm-unk_template-train_200-test_159-valid_100-SD_0, templates never seen in training, 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)	59.6% (53.7 to 65.3)	3.1% (0.6 to 6.3)	7.6% (5.4 to 10.3)	1.9% (0.0 to 4.4)	US$0.19 (0.17 to 0.20)	159
C, Gemma 4 E4B untuned on the image	47.2% (42.6 to 51.9)	0.0% (0.0 to 0.0)	13.5% (10.4 to 16.9)	0.6% (0.0 to 1.9)	US$0.22 (0.20 to 0.24)	159
A, Claude Sonnet 5.5 on the image	69.8% (65.4 to 74.2)	0.6% (0.0 to 1.9)	10.0% (7.7 to 12.4)	0.0% (0.0 to 0.0)	US$5.19 (4.94 to 5.45)	159
A, Claude Haiku 5.5 on the image	66.4% (61.5 to 71.0)	1.3% (0.0 to 3.1)	11.1% (8.7 to 13.8)	0.6% (0.0 to 1.9)	US$0.35 (0.33 to 0.37)	159
A, GPT-6.1 Sol on the image	71.4% (66.7 to 76.0)	2.5% (0.6 to 5.0)	8.6% (6.5 to 10.9)	3.8% (1.3 to 6.9)	US$4.53 (4.32 to 4.75)	159
A, Gemini 3.8 Flash on the image	70.0% (64.4 to 75.1)	2.5% (0.6 to 5.0)	7.0% (5.3 to 9.0)	3.1% (0.6 to 6.3)	US$2.73 (2.39 to 3.14)	159

The best frontier model on these documents is GPT-6.1 Sol, at 71.4% (66.7 to 76.0) of fields for US $4.53 (4.32 t o 4.75) a t h o u s an d p a g es . T hi s a d a pt e r' s f i e l d a cc u r a cy d i f f er s f r o mi t b y - 11.8 p o in t s, a t U S$ 0.19 (0.17 to 0.20) a thousand pages, about 1/24 of the cost. Against its untuned base it reads +12.4 points.

Training variance: the same recipe at seeds 1 and 2, each on its own split's test set: seed 1, 8.0% (4.6 to 12.1); seed 2, 57.1% (51.1 to 62.9). The three seeds' point estimates span 51.6 points (their test sets differ where the split files do).

The 03 release gate

Against Claude Haiku 5.5: block (fields pass, block, under-powered: 0, 9, 5); worst: line_item-channel -29.6% (-38.8 to -19.7); contract_num -23.3% (-30.2 to -15.7).
Against GPT-6.1 Sol: block (fields pass, block, under-powered: 0, 10, 4); worst: line_item-channel -32.2% (-41.8 to -22.6); line_item-sub_amount -25.0% (-33.6 to -16.4).

Failure modes

A form design it has not seen. The README's two tables, templates seen in training and not, for this size and the best frontier model: Gemma 4 E4B 58.2% (53.0 to 63.2) on templates seen in training and 59.6% (53.7 to 65.3) on templates not seen (+1.4 points); GPT-6.1 Sol 72.3% (68.5 to 75.9) on templates seen in training and 71.4% (66.7 to 76.0) on templates not seen (-0.9 points).
A bad scan. On the 30 of its test documents in the VRDU curve subset, real scans, degraded in the project's five reproducible steps (docs/degradation.md): level 0 67.7% (53.3 to 79.1), level 2 64.0% (49.2 to 75.5), level 4 63.0% (47.8 to 75.6).
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 (2,276), 81.7% (72.2 to 89.4) are right; the commonest error there is a wrong value (193). At 99% accuracy it sends 99.8% (99.5 to 99.9) of its test fields to a person. 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 8.0% to 59.6% 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 blocks 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. 1.9% (0.0 to 4.4) 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. 7.6% (5.4 to 10.3) 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 DeepForm ad-buy forms 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