SN91, Cascade

Training time series foundation models
On Bittensor Subnet 91
By @tensorlink_dev

[SN91, Cascade @cascade_sn91]

This week on Cascade:

Four releases moved the live competition forward.

✓ Warm starts are now live. The winning checkpoint now carries forward for five rounds (roughly 2.5 days) before a challenger can replace it.

✓ Signed public benchmark reports are now generated for champion and challenger after each round, creating a tamper-evident performance record.

✓ Fleet-wide hardware telemetry now records data from every worker, helping Cascade measure whether machine differences influence results.

✓ Safeguards around reporting, restarts and GPU rentals were strengthened, keeping promotion cycles moving and validators aligned.

More to come next week.

[SN91, Cascade @cascade_sn91]

@ukwealthbee Yes we have some announcements to come soon on product roadmap and a stakeholder dashboard to track progress of milestones

[SN91, Cascade @cascade_sn91]

Cascade just moved from cold starts to warm starts - the next step in the competition’s design.

Until now, every competition began with a fresh Toto-2 4M model, trained from random initialisation. That gave us a controlled way to answer the first question:

Which miner’s synthetic data produces the best forecaster?

Now, the strongest checkpoint produced by the competition becomes the shared starting point for the next phase.

Every generator still begins from the same model, keeping the comparison controlled. But instead of rebuilding from zero, the network can improve the strongest forecaster it has already produced.

Cold starts find the best data. Warm starts compound what it teaches the model.

[SN91, Cascade @cascade_sn91]

@ukwealthbee Warm start allows us to continue training model "to completion". If we use the industry standard benchmarks people just overfit to benchmark (we generate fresh evals daily), but now we are moving to warm start we will begin reporting (but not scoring) on industry benchmarks.

[bitstarter @bitstarterAI]

https://t.co/TEF2mVFiG8

[SN91, Cascade @cascade_sn91]

@LordRagnarao @ukwealthbee See our response here: https://t.co/DxrMVzMxIE

[SN91, Cascade @cascade_sn91]

Since last week's Novelty search there have been serious allegations against Quasar, backed by evidence, that their model was not pretrained by them. The team stands by their work and says they'll post open weights on Monday. We'll judge that evidence like everyone else.

Whatever Monday shows, this is extremely disappointing. Bittensor runs on a huge amount of goodwill. Investors backing subnet teams, operators trusting each other, valis trusting miners. Incidents like this create distrust across the whole network, and the people who actually pay for it are stakeholders and, more than anyone, sn24 token holders.

On Cascade we try to make trust unnecessary. Our incentive mechanism and results are set up so that anyone can audit them, with multiple tiers of checks on every result, and all model checkpoints public as they are trained.

For the record: Cascade is incubated through the Bitstarter ML track but has no direct
affiliation with sn24.

[SN91, Cascade @cascade_sn91]

Since last week's Novelty search there have been serious allegations against Quasar, backed by evidence, that their model was not pretrained by them. The team stands by their work and says they'll post open weights on Monday. We'll judge that evidence like everyone else.

Whatever Monday shows, this is extremely disappointing. Bittensor runs on a huge amount of goodwill. Investors backing subnet teams, operators trusting each other, valis trusting miners. Incidents like this create distrust across the whole network, and the people who actually pay for it are stakeholders and, more than anyone, sn24 token holders.

On Cascade we try to make trust unnecessary. Our incentive mechanism and results are set up so that anyone can audit them, with multiple tiers of checks on every result, and all model checkpoints public as they are trained.

For the record: Cascade is incubated through the Bitstarter ML track but has no direct
affiliation with sn24.

[Chris(t)ensor @tensorlink_dev]

3 dethrones, 4 theories of what data teaches a forecaster best, settled by head-to-head evals nobody can overfit. The last four duels all landed within ±5%.
@cascade_sn91

Fetch any king's generator yourself: `cascade fetch king` 👑

[SN91, Cascade @cascade_sn91]

https://t.co/DeYA3yRuwh

[Chris(t)ensor @tensorlink_dev]

https://t.co/n6NkoewMEG

[SN91, Cascade @cascade_sn91]

Extrinsic here

https://t.co/HWib1Uknpb

More updates from the team via our Discord

https://t.co/btUuavSSBX

[SN91, Cascade @cascade_sn91]

Confirmed ☑️ SN91 enters mainnet with Conviction

50% owner alpha
80,000 tokens
$230k market value

All-in on training time series models. All-in on #SN91.

[SN91, Cascade @cascade_sn91]

@taodotcom @bitstarterAI More spam protection, less burn fees for miners

[TAO.com @taodotcom]

UPDATE: @bitstarterAI’s Cascade Subnet 91 reduced min_burn from 200,000,000 to 100,000,000 units on July 27. The on-chain change halves the minimum burn required to register a new miner hotkey.

[SN91, Cascade @cascade_sn91]

View the dashboard: https://t.co/VDqpoqhBs3
Docs on mining Cascade: https://t.co/gAYoPCrz8A
The code's public if you think you can build better training data. 📈

[SN91, Cascade @cascade_sn91]

Most training subnets have miners submit model weights. Cascade #sn91 doesn't.

Every submission trains the same 4M-parameter model from scratch, with the same seeds and the same compute budget. The only thing a miner submits is a data generator.

So the thing you compete on is the training data.

Miner spotlight 🧵

[SN91, Cascade @cascade_sn91]

Each earlier version ties a change in the weights to a weakness measured on real data. Weak on financial series, so more slow-drifting processes. Weak on economic data, so more regime changes. Retail, so more demand spikes.

They're tuning training data like a hyperparameter.