turion-ai

turion-extract-alpha

README

License: apache-2.0

Request and answer

python

from turion_extract.predict import Extractor   # pip install "./code" ext = Extractor.from_dir("turion-extract-alpha")   # a local copy of this repo (see below)ext.extract({    "state": "Invoice INV-2291 from Northwind Supply, issued 2026-09-02. 4 pallets at $250.00, shipping $80.00. Contacts: Dana Ruiz, Priya Shah.",    "fields": {        "invoice_number": {"type": "text"},        "total": {"type": "number", "decimals": 2, "instructions": "Amount due including shipping"},        "contacts": {"type": "list", "items": "text"},        "po_number": {"type": "text", "instructions": "Purchase order number"},    },})# {"answers": {"invoice_number": {"type": "text", "value": "INV-2291", "span": [8, 16], ...},#              "po_number": {"type": "text", "value": null, "p_null": 0.98, ...}, ...}}

Extractor.from_dir takes a local directory; download this repo first (hf download turion-ai/turion-extract-alpha --local-dir turion-extract-alpha). The base model downloads on first use.

How it works

Layout. The field list comes before the document, so every document token is read knowing the fields, and one query line per field comes after it.
Heads on the last hidden states:
- a mode per field (null, span, sum, count, pick, difference, list), which gives p_null;
- start/end pointers over the document for text;
- for numbers, a sign (0, +1, −1) for every number a parser finds in the text, so sums and differences come out exact, plus a count head;
- a pick over the dates and times the parser found (both readings of 03/10/2026, relative dates against reference_date);
- a difference of two dates or times for durations;
- begin/inside tags for lists.
Values are built by plain code from what the heads point at, so every answer is well formed.
Loss: −log of the total probability of every derivation that gives the gold value.

Training

LoRA r32 (alpha 64) on all linear layers of the language model, plus the heads.
Learning rates: LoRA 1e-4, heads 1e-3.
299 steps of ~65k tokens: one pass over 39,104 documents (19.5M tokens).
3 h 22 min on one RTX PRO 6000, as a Turion AI platform job.
Data (training splits only):
- SQuAD 2.0 (CC BY-SA 4.0);
- DROP (CC BY-SA 4.0);
- TAT-QA (CC BY 4.0);
- CUAD (CC BY 4.0);
- CORD v2 (CC BY 4.0);
- Few-NERD (CC BY-SA 4.0);
- 15,000 synthetic business documents from our own generator: 32 kinds, 27 locales, exact labels. It was written independently of the index's docs generator.

Limitations

Computation: sums and differences of up to 3 numbers in the text, counts, and durations between two dates or times. Not yet: ratios, products, percentages, averages, or dates computed from offsets ("issued plus net 30"). Most of TAT-QA's and DROP's misses are these.
Weak types: date-times and durations in unfamiliar layouts, such as the index's docs test.
Length: up to 16,384 tokens per request; up to 24 fields per pass, with more split across passes.
Input: text only; no images yet.
Probabilities: p_null and confidence are not calibrated beyond the training mix.

Licence

Apache-2.0, like the base model; see LICENSE and NOTICE. The training sets keep their own licences, listed above.

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

turion-ai

Model Tree

Base

google/gemma-4-12B-it

Adapter

this model

Input Modalities

TextAudioImageVideo

Output Modalities

Text

Supported Functionality