MIST-Extract-1.0
Documents in, JSON out, with a confidence on every field.
MIST-Extract-1.0 turns a document into structured data. Give it a JSON template and a document as text, a photo or scan, or both, and it fills in the template, telling you how likely each value is to be right. It is a 4B model built for the everyday documents of African businesses and their customers, in English and African languages.
The problem it solves
A large share of business data still arrives as documents: a photo of a market receipt, a mobile-money SMS, a utility bill, a tenancy letter. General-purpose extractors are built on Western invoices and forms, and they stumble on naira and cedi amounts, local layouts and names in African languages.
MIST-Extract was trained on exactly those documents, and it does not just return values: every field carries a calibrated confidence, so you can keep the certain fields automatic and send only the doubtful ones to a person.
Highlights
Whole documents right, not just fields
On 486 held-out English documents (receipts, notices, emails, contracts and more) it gets 91.8% completely right. NuExtract3, built on the same 4B base, gets 67.9%. On English receipt photos it gets every field right.
Strongest on African languages
On MasakhaNER 2.0 named entities in 21 African languages it scores 0.728 field accuracy, against 0.393 for NuExtract3, 0.432 for Qwen3.5-4B and 0.452 for GLiNER2.5-multi.
Confidence you can threshold
When it says 0.66, it is right about 66% of the time: calibration error of 0.008 on held-out fields. Values it rates 0.99 and above are right 99.9% of the time.
Drop-in templates, text or image
Uses the same template format as NuExtract, so existing templates move across unchanged. Reads text, one image, or both, and served at a 131,072-token context.
Run it in a few lines
import sys
from huggingface_hub import snapshot_download
path = snapshot_download("olaverse/MIST-Extract-1.0")
sys.path.insert(0, path) # the repo ships mist_extract.py
from mist_extract import Extractor
x = Extractor.from_pretrained(path, device="cuda")
template = {
"merchant": "verbatim-string",
"date": "date",
"total": "number",
"payment_method": ["cash", "card", "mobile money", "bank transfer"],
"items": [{"name": "verbatim-string", "quantity": "integer", "price": "number"}],
}
out = x.extract(template, image="receipt.jpg") # a photo or scan
out = x.extract(template, text="Paid 2,500 naira at Mama Put on 3 May 2024 with my card. 2 plates of jollof.")
out["data"] # {"merchant": "Mama Put", "date": "2024-05-03", "total": 2500, ...}
out["confidence"] # same shape, a number in [0, 1] for each fieldWhich MIST-Extract model?
Need more throughput? MIST-Extract-Flash-1.0 takes the same templates at 2B and runs 1.6× faster.
| MIST-Extract-1.0THIS PAGE | MIST-Extract-Flash-1.0 → | |
|---|---|---|
| Parameters | 4B | 2B |
| English documents fully right (held-out) | 91.8% | 90.5% |
| Named entities, 21 African languages | 0.728 | 0.701 |
| Field accuracy, all held-out documents | 0.898 | 0.879 |
| Speed, one RTX A6000 | 6.0 docs/s | 9.7 docs/s |
| Confidence calibration error (ECE) | 0.008 | 0.022 |
| Best for | Top accuracy, general and African documents | Volume and receipts |
Best for
- Mobile-money and bank messages: transfer confirmations, wallet SMS, alerts and statements.
- Receipts and bills in naira, cedis, shillings, rand and other African currencies, from text, photos or screenshots.
- Letters, forms, IDs and tickets, with names of people, places and organisations in 20+ African languages.
- Review queues: automate the confident fields and route only the uncertain ones to a person.
Use cases
Document extraction
Fill a JSON template from a photo, a scan or plain text, with a calibrated confidence on every field. Built for African documents and currencies, in English and African languages.
/ Extraction · DecisionsPayment messages
Pull the amount, sender, reference and date out of mobile-money and bank SMS, then decide in one pass whether it is genuine, which account it belongs to and whether anyone needs to look at it.
More production models
All modelsPurpleMIST-Flash-1.0
typed-decisions 0.720 · African TD 0.668
PurpleMISTPurpleMIST-Mini-1.0
typed-decisions 0.601 · 1.9B
LID-608lid-neural-608
0.867 acc · 140M
LID-608lid-lite-608
37 MB · 6,800 texts/s
DiacNetdiacnet-2.0
DER 0.0117 · 582M
DiacNetdiacnet-mini-2.0
DER 0.0141 · 300M
DiacNetdiactag-2.0
DER 0.0130 · 37.9M
MIST QGmist-qg-1.5b
0.955 keep-rate · 25 langs