Public Training Report: Dainvo T-JEPA Scheduler Model

Report date: June 12, 2026

Summary

This report explains the latest training work for the Dainvo T-JEPA scheduler model.

Dainvo T-JEPA is a schedule and workflow representation model. It is not a full scheduling assistant by itself. Its job is to learn useful patterns from schedule like data, such as meetings, tasks, deadlines, email activity, project work, public calendar events, and planning examples.

The latest work did two things:

1. It added a new group of public datasets related to calendars, planning, tool use, task planning, meeting scheduling, and agent workflows.
2. It ran a new round of model training across all prepared public datasets, first in a fixed order and then in randomized dataset orders.

The final training pool contained:

All latest runs completed successfully. Loss values stayed finite, and every planned checkpoint was saved.

Earlier Reports Referenced

This report builds on two earlier reports.

Earlier Report 1: June 11 T-JEPA Training Report

The June 11 report explained the first full public-data training run. It described:

• the JEPA learning objective;
• the 96 slot daily window format;
• the 16 numeric features used for each time slot;
• the first set of public datasets;
• six randomized confirmation passes;
• a baseline comparison against an untrained model.

That earlier baseline comparison showed that the trained model had much lower validation loss than a fresh untrained model on the same objective. That result did not prove real world scheduling quality; it showed that the representation model was learning from the prepared data.

Earlier Report 2: June 12 Hugging Face Data Expansion Report

The second report covered the new data expansion. It added public datasets for:

• calendar and meeting scheduling;
• trip and natural language planning;
• tool use planning;
• task planning;
• agent planning;
• shift and staff scheduling;
• workflow style instruction data.

Those sources were converted into one prepared dataset group called Expanded scheduling, planning, and agent data.

That new group produced:

What The Model Learns

The model learns from partial schedule-like windows.

Each training example is a fixed size grid:

96 time slots x 16 features

The 96 slots represent a day at about 15 minute resolution. Each slot has 16 simple numeric features, such as:

• whether the slot looks like a calendar event;
• whether it contains a task;
• whether it relates to email activity;
• whether it relates to project work;
• whether it has a deadline;
• how long the event or task may last;
• whether it appears high priority;
• how many people may be involved;
• how much text or description was attached.

The model is trained by hiding part of the window and asking it to predict the hidden part in representation space. This is self supervised learning. It does not require a person to label each example as good or bad.

The intended benefit is that the model learns common structure across many types of scheduling and workflow data.

Data Used In The Latest Training Runs

The latest training used 13 prepared public dataset groups.

Latest Training Plan

The latest training had two stages.

Stage 1: Ordered Full-Data Run

The model first trained through every prepared public dataset in a fixed order.

This gave a clean reference run where the data order was predictable.

Stage 2: Randomized Mixed Runs

After the ordered run, the model trained through the datasets again in randomized orders.

There were three randomized passes. Each pass used the same full prepared data pool, but the dataset order changed. Each pass also trained a combined model over the full mixed data stream.

This was done to check that the model could train consistently when the source order changed.

Training Settings

The latest runs used the same core settings:

Streaming means the training code read batches from prepared files as needed instead of loading the entire multi million-window dataset into memory at once.

Combined Model Results

The combined model is the most important result because it trains over all public prepared dataset groups together.

These values are all low and finite. That means the training process was stable for the ordered run and for the randomized runs.

The validation loss changed somewhat between randomized passes, which is expected. The training order changed, and each run starts a new model training pass rather than continuing one single model forever.

Ordered Run: Source by Source Results

The ordered run trained through the dataset groups in a fixed sequence. This was the clean reference pass before randomizing dataset order.

Final Randomized Pass: Source-by-Source Results

The final randomized pass is useful because it was the last complete shuffled run.

Most datasets ended with low validation loss. The Calendar group was harder for the model than the other groups. That is not automatically bad. It may mean that this source has more varied language, noisier structure, or patterns that do not map as cleanly into the current 16 feature schedule format.

What The Results Mean

The latest runs show that:

• the training pipeline can handle all prepared public datasets together;
• the new expansion was successfully included;
• the model trains stably on millions of windows using streaming;
• randomized dataset order did not break training;
• the combined model kept low validation loss across the ordered and randomized runs.

In simple terms: the model is learning the internal schedule representation task across a much broader collection of public scheduling, planning, project, email, and workflow data than before.

What The Results Do Not Prove Yet

These results do not yet prove that the model makes better real world schedules.

The current loss measures whether the model learned to predict hidden parts of prepared schedule like windows. That is an important representation learning test, but it is not the same as asking:

• Did the model choose the best meeting time?
• Did it reduce conflicts?
• Did it improve task prioritization?
• Did it produce schedules people prefer?

Those questions require a downstream scheduler evaluation.

Data and License Notes

The training data was drawn from public sources collected for the project. Some public sources have clear permissive licenses, while others require extra review before external release or commercial reuse.

For public communication, the safest statement is:

The model was trained on a mixture of public scheduling, planning, workflow, project, email-derived, and calendar-style datasets. Some sources require license review before redistribution or commercial use.

No private user calendar exports, private tenant data, or personal mailbox exports were included in this training pass.

Main Takeaways

1. The model now trains on 13 public prepared dataset groups instead of the earlier 11.
2. The new expansion set added more planning, scheduling, and agent style examples.
3. The latest full training pool contains about 6.8 million training windows and 754 thousand validation windows.
4. The ordered full data run completed successfully.
5. Three randomized mixed passes completed successfully.
6. The final combined randomized pass ended with validation loss of 0.0125.
7. The training results support continued work, but downstream scheduling tests are still needed.