Post-training, fully automated

Forge trains an AI you own to be the best at your business's work. It learns from your real conversations, keeps every model that does the job better, and the model is yours to keep.

Open the lab See the example

TrafficFind mistakesAgentsTrainResult

Your agent's real conversations come in

Built with

See it learn

Questions it used to miss, it now gets right

A question from the final test

The Ben and Aggie Burrito Shop makes 125 chimichangas on Tuesdays, 125 chimichangas on Wednesdays and twice as many on Friday. How many chimichangas do they make on those three days?

Before ForgeThe original model

To determine how many chimichangas the Ben and Aggie Burrito Shop makes on each of the three days (Tuesday, Wednesday, and Friday), we can follow these steps:

1. Identify the number of chimichangas made on Tuesday.
   - The shop makes 125 chimichangas on Tuesday.

2. Identify the number of chimichangas made on Wednesday.
   - The shop makes twice as many chimichangas on Wednesday as it does on Tuesday.
   - Therefore, the number of chimichangas made on Wednesday is \(2 \times 125 = 250\).

3. Identify the number of chimichangas made on Friday.
   - The shop makes 125 chimichangas on Friday.

4. Sum up the total number of chimichangas made over the three days.
   - Total chimichangas = Chimbachamas on Tuesday + Chimbachamas on Wednesday + Chimbachamas on Friday
   - Total chimichangas = 125 + 250 + 125

Now, let's perform the addition:
\[125 + 250 = 375\]
\[

WrongMisreads the question, then stops before giving an answer.

After ForgeTrained by Forge

On Tuesday, they make 125.
On Wednesday, they make 125.
On Friday, they make twice as many as on Tuesday, which is 2*125 = <<125*2=250>>250.
In total, they make 125+125+250 = <<125+125+250=500>>500 chimichangas.
#### 500

CorrectReads each day right and ends with the answer: 500.

Both answers exactly as the models wrote them. Highlights mark where each one goes wrong or right.

How Forge works

From your conversations to the best model for your work

It finds the mistakes

Upload your agent's conversations. Forge spots the answers that need fixing, all by itself.

It checks every answer

Missing answers, replies that were cut off, math slips and corrections from users.

Its flags hold up

Every answer it flagged was confirmed by a person.

The final test stays private

Anything that matches the final test is set aside first.

Answers Forge flaggedFrom your agent's traffic

Willie came back from Europe with 70 euros. Normally 5 euros is worth 1…Missing answerCut off
The balance of Pam's bank account tripled during the year. At the end of the…Math slipMissing answerCut off
Fred was having a party and was responsible for buying canned soda. He…Missing answerCut off
A math teacher had $100 to buy three different types of calculators. A basic…Missing answer
Robbie weighs 100 pounds. Patty was 4.5 times as heavy as Robbie, but she has…Missing answerCut off

Five agents do the work

A researcher, a director, a trainer, a tester and a council that approves. They talk to each other, and every message is kept.

Direct messages

Each handoff goes to one agent, who confirms it before starting.

A shared room

Plans and results are posted for the whole team.

Replies in threads

The council answers right under each new version.

Read receipts

You can see who read each message.

researcher

director

engineer

evaluator

council

Data researcherResearch directorDirect message
I found the answers that need fixing. Half will teach the model, and half are saved to test it.
Read by the director

A record you can always resume from

Every goal, approval and saved file lives in a permanent record, so a run picks up exactly where it was.

Stopped on purpose

We force-stopped Forge in the middle of training.

Back in seconds

Every record came back exactly as it was.

Right where it left off

Training carried on from the very same step.

A test we ran during trainingRecovering

Stopped on purpose

We force-stopped it mid-training

Restarting

The runtime comes back up

Records restored

Everything exactly as it was

Training resumes

From the very same step

Goal and budgetrestored
Experimentsrestored
Approvalsrestored
Saved filesrestored

A model that is yours

Forge tests the new model on brand-new questions. Then you download it and run it wherever you like.

Tested on new questions

Some real mistakes are kept aside to check the model after training.

A fair final exam

The final test opens only after the best version is chosen.

The best version wins

The council promotes the version that did best in practice.

Yours to keep

Download the model, the report and every step of the research.

See the example

From mistakes to your model

TrainOn the answers it missed
TestOn brand-new questions
ApproveThe governance council signs off
DownloadYour model, your report, every step

Under the hood

Every step leaves a trail you can check

It learns from your traffic

Your agent's own answers become the lesson plan.

The team checks its own work

The trainer hands each new version to the tester, and the council reviews it before anything is promoted.

Every step is signed and saved

Each change is signed and written down, so the whole run can be replayed.

Results

Better at your work, and it keeps improving

Fixes what it missed

After training, it answers questions it used to miss.

Keeps what it knew

Everything it already answered well, it still answers well.

Measured fairly

It is tested on brand-new questions, held back from training.

Yours to run

You download the model and own it, along with what it costs to run.

The record

Every decision is on the record

Each message, review and saved file is kept with the run, so you can trace every result back to where it came from.

Versions

First versionPromoted by the council
Second versionKept on the record

Every version, with its evidence

Each version keeps its settings, its results and the council's decision.

Council reviews

Plan checkApproved before training
Version reviewApproved after testing
Plan checkApproved before training
Version reviewApproved after testing

A signed approval for each step

Every review is tied to a written policy.

Saved files

Where the data came fromSaved at the start
The model before trainingTested and saved
The new modelSaved after training
Final test resultsSaved at the end
The full reportReady to download

Everything is downloadable

The model, the report and the full research trail.

Agent Relay

The agents talk to each other

What the agents said

I found the answers that need fixing. Half will teach the model, and half are saved to test it.

DRData researcher to Research directorDirect message

Messages from a real Forge run, in plain words.

What the run remembers

Six records, each kept by its own durable actor and saved as the run goes, so a crash picks up where it left off.

Goal and budgetWhat the run is for and what it may spend

ExperimentsEvery setting and every result

ApprovalsA signed decision for each new version

Saved filesEvery file, with a fingerprint to check it

Practice problemsThe worlds that score each answer

Research notesEvery idea and decision, in order

The lab

Watch every run as it happens

The agents' conversation, every version and the results live in one place while the run is going.

The Forge lab: the five agents' conversation in the research room

Your work.
Your model.