weather-rescue-brazil-reader

Weather Rescue Brazil — row reader

A small open model that reads one printed table row out of a scanned nineteenth-century meteorological table and returns its cells as text.

It is not a chatbot, not a general OCR model, and not a climate model. It reads a strip of paper and writes back numbers.

input : a crop of one row from a printed table
output: 57.23 | 58.08 | 55.67 | 22.37 | 23.2 | 21.3 | 16.75 | 83.4 | C, SSE | 3.3 | 7.4 | 2 | 9

Trained and evaluated entirely on Brazilian historical records, but the interesting result is that it reads layouts it has never seen — and usually with no labelled rows at all. Of the 18 layouts this project describes, two ever needed hand-labelling; the other fifteen were added by writing a JSON description of the columns and running the pipeline. docs/ADDING-A-PUBLICATION.md is the path and the evidence.

Base and adapters

Base: Qwen/Qwen3.5-2B, with LoRA (r=16) on the language layers only. Training ran on a Mac (MPS), offline, ~45 minutes per epoch.

Three adapters are published, and it matters which you take:

gen3 is the one that produced the dataset in the companion repository, and it is deliberately not the highest-scoring adapter here: it was trained to keep reading when the layout changes, which costs a little on the layouts it knows. Report the adapter you actually measured.

Measured accuracy

On a frozen, triple-transcribed gold set of nine pages (273 rows, 3,822 cells), never trained on:

Blind tests on layouts it had never seen, with gen3:

Added 2026-09-15, on layouts absent from the corpus when the adapter was trained and given no labelled rows at all:

The useful correction in those three: a claim that the model could not read rows wider than 16 columns had shelved a whole layout as needing a trained variant. It reads 18 columns in one pass and 26 in two bands. What had failed was the crop.

Added 2026-09-18, the fourth Radcliffe table and the first English-language layout added with no labelled rows and no new code — only a JSON file copied from its neighbour on the same sheet:

What that table cost was writing down its columns. What the barometer table on the facing page cost was three separate printing conventions the pipeline had never met — an integer part elided down the column, a raised decimal point, and a double rule the reader takes for an empty column — and it still only reaches 15 of 25. The lesson is the one this project keeps relearning: the model reads what it is shown, and the work is in what you show it.

Training-data provenance — please read

The training labels were produced by the Gemini API, not by hand. The photographs were read by a frontier model and its output, after human review of disagreements, was used as the training target for this adapter. This is declared because it matters: the adapter inherits the label generator's conventions, and anyone redistributing or building on these weights should know where the targets came from.

The evaluation gold set, by contrast, is human — triple-transcribed, with every disagreement adjudicated against the image.

Known limits

• It counts cells from training, not from the image. A layout whose printed column count it has not been balanced on will come back with the wrong number of cells. Balancing distinct examples per convention is the fix, and it is documented; stating the count in the prompt does not work.
• Two values in one printed cell defeat it. The 1883 nébulosité and wind tables print a number and a cloud-form code stacked in the same column, and the model's cell count is unreliable there.
• A ~1% residual of consistent, in-range misreads. These are the same wrong digit on every read, so neither voting nor re-reading at another scale finds them. Only a human comparing against the page does. Three automatic detection strategies were tried and all failed.
• It does not read a whole page. Row localisation is done by geometry in the companion repository; this model reads one prepared crop at a time.

Using it

from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel

base = "Qwen/Qwen3.5-2B"
model = AutoModelForImageTextToText.from_pretrained(base, dtype="bfloat16")
model = PeftModel.from_pretrained(model, "<this repo>/gen3").eval()

The prompt is fixed and lives in the companion repository:

Transcreva esta linha de tabela impressa exatamente como está, célula por célula, da esquerda para a direita, separando as células com ' | '. Inclua todas as colunas, inclusive número do dia e texto. Use 'null' para célula vazia. Só a lista.

Read the row as printed. The publication's conventions — an elided leading digit on a barometer, a ditto mark meaning "the same as the row above" — are undone in code afterwards, not by the model. Teaching the model the reconstructed form taught it to invent digits.

Licence

Apache-2.0 — the same terms as the base model it adapts (Qwen3.5-2B), which this does not alter. Commercial use is permitted.

Until 2026-09-17 these weights were CC BY-NC 4.0. That was changed so the method could be adopted by the people most likely to want it — national meteorological institutes with their own shelves of scanned tables — and so the work could go to venues that require an OSI-conforming licence. Read the provenance section above before building on it: the training labels came from the Gemini API. Google claims no ownership of generated content, and its use restriction on developing competing services binds the API customer rather than travelling with these weights, but anyone redistributing them should know where the targets came from.

Companion

Pipeline, profiles, the dataset and every measurement above: github.com/akagabi/weather-rescue-brazil, archived at 10.5281/zenodo.22876699.