Orkhon's first model

Bengü-1

An email decision model. It reads an incoming email and returns a typed decision: a category and an action, with a probability. It does not write text.

In use in-house. Not publicly released.

The name

bengü

Old Turkic: eternal.

The Orkhon stones call themselves bengü taş, the eternal stone.

It decides. It does not write.

Bengü-1 reads one incoming email and returns a typed decision: a category and an action, with a probability. It does not draft, summarize or reply.

One incoming email goes in. Bengü-1 returns a typed decision with a category, an action and a probability. It writes no text.
Model Bengü-1
Out A typed decision
category
returned
action
returned
probability
returned
text
never
The decision as a type: Bengü-1 maps an email to a decision with a category, an action and a probability, and no text field.
bengü-1 : Email -> Decision

Decision = {
  category    : Category,
  action      : Action,
  probability : Probability
}

// no text field

Built on laya-multilingual.

Bengü-1 starts from convaiinnovations/laya-multilingual, a base model released under the Apache 2.0 licence. We credit it on the model card.

Base model
convaiinnovations/laya-multilingual
Licence
Apache 2.0
Size
About 322M parameters
Credit
On the Bengü-1 model card

Real business email. It never leaves our machine.

Bengü-1's training data is real business email. A local model labels it, and the mail never leaves the lab's own machine. Some mail never enters the data set at all.

The lab's own machine

Never enters the data set

  • Passwords and codes
  • Health
  • Family and personal mail
  • Legal mail
  • Bank statements

Everything else is scrubbed

Details are replaced with typed placeholders such as [PERSON] and [AMOUNT].

Labeled by a local model

The labels come from a model on the same machine. No mail is sent out to be labeled.

Leakage audit on a sample

A sample of the scrubbed mail is checked for anything the placeholders missed.

The mail stays inside this frame: on the lab's own machine.

A gold set, held out and split by time.

Bengü-1 will be evaluated on a held-out gold set, split by time: tested on later mail than the mail it learned from.

EarlierTraining mail
LaterGold set, held out
Schematic, not to scale.

When we release Bengü-1 publicly, we will publish a benchmark on a public data set against existing decision models and the base model, including the cases where we lose.