cascade: SOTA time-series foundation models on Bittensor

cascade is building state-of-the-art time-series foundation models (TSFM) on Bittensor. We start where the leverage is: data. The first task holds the model byte-identical so the only variable is data quality, and scores the data generators that feed it — better synthetic data, better forecasters. The full sequence is in the technical roadmap.

Why compete on data

Synthetic data isn't cascade's only lever, but we believe high-quality synthetic data is critical to training a time-series foundation model. That view tracks the consensus direction of the field: recent models keep winning on benchmarks by competing on synthetic priors, not architecture.

• Chronos-2 (Amazon, 120M) reaches state-of-the-art zero-shot accuracy on fev-bench, GIFT-Eval, and Chronos Benchmark II, trained heavily on large-scale synthetic series (Gaussian-process curves, trend/seasonality/ irregularity mixtures, random temporal causal graphs) (arXiv 2510.15821). Their ablation trained purely on synthetic data, Chronos-2-Synth, stays within ~1 skill point of the full model on GIFT-Eval (50.4 vs 51.4) and Chronos Benchmark II (46.4 vs 46.6) — the authors note this suggests real data "may not even be required for effective pretraining."
• FlowState (IBM, 9.1M) is the smallest model in GIFT-Eval's top 10, out-forecasting rivals 20x+ its size, pretrained in part on synthetic series from the CauKer generator (arXiv 2508.05287).
• ForecastPFN is a prior-data fitted network trained purely on a synthetic distribution, and was the first zero-shot forecaster to beat the then-SOTA with no real training data at all (arXiv 2311.01933); TempoPFN (arXiv 2510.25502) carries the purely-synthetic pretraining recipe further.
• DynaMix (NeurIPS 2025) is trained on nothing but a narrow synthetic corpus of 34 chaotic dynamical systems and, with ~0.1% of Chronos's parameters (~10k in total), still beats Chronos zero-shot on real-world traffic and weather it never saw: a small, well-curated synthetic prior outperforming a far larger real-data model (arXiv 2505.13192).

The throughline: across the leaderboard, the synthetic data distribution is doing the heavy lifting. cascade turns that distribution into the competitive surface, holding the model fixed so miners compete the prior.

How it works

The fixed process is a Toto2-4M backbone trained from random initialisation (Datadog/Toto-2.0-4m, arXiv 2605.20119), not a fine-tune of released weights. Training from scratch is the point: the corpus is then the only source of learned signal, so the downstream forecast skill measures the data, not what some pretrained checkpoint already knew. Toto 2.0 itself is 57.5% synthetic data with zero public series in pretraining and still tops GIFT-Eval, and cascade turns that synthetic-prior design into an open competition.

flowchart TD
    subgraph miner["Miner (no GPU)"]
        gen["generator.py<br/>(DataGenerator)"]
        upload["push to Hippius Hub registry<br/>(OCI) → repo@digest"]
        commit["commit on-chain pointer<br/>metro-v1:gen:hippius:repo@digest"]
        gen --> upload --> commit
    end

    subgraph trainer["Trainer: owner-operated (the GPU boundary)"]
        resolve["resolve commitments before the<br/>24h epoch cutoff → king + field"]
        seeds["derive one shared RoundSeeds<br/>from epoch-boundary block hash<br/>(generation_seed + training_seed)"]
        heat["HEAT: train every challenger<br/>~30min (primary size) → screen<br/>→ top finalist"]
        trainK["FINAL: train king + finalist<br/>from random init at EVERY size<br/>(Toto2-4M, Toto2-22M)"]
        upK["push ckpts → Hippius Hub registry<br/>logs/metrics → Hippius S3"]
        manifest["sign + publish TrainingManifest<br/>to Hippius S3 (size-tagged ckpt refs + digests)"]
        resolve --> seeds --> heat --> trainK --> upK --> manifest
    end

    subgraph validator["Validator (eval GPU)"]
        gate["verify signature +<br/>matching contract / base-arch digests<br/>(controlled-experiment gate)"]
        eval["pull king + finalist ckpts per size →<br/>score on shared held-out windows<br/>(CRPS/MWSQL + MASE)"]
        koth["paired-bootstrap LCB of<br/>geomean(CRPS, MASE) POOLED across sizes,<br/>finalist vs king → one KOTH verdict"]
        weights["equal-share weights<br/>(king + recent kings)"]
        gate --> eval --> koth --> weights
    end

    commit -->|on-chain| resolve
    manifest -->|S3 manifest + registry ckpts| gate

    classDef invariant fill:#fff3cd,stroke:#d39e00,color:#5c4400;
    class seeds,gate invariant;

Loading

The highlighted boxes are where the controlled experiment lives: the trainer reuses one RoundSeeds for every run in the round, and the validator's digest gate rejects any manifest where king and challenger didn't share that contract. Details below.

The round cadence. A round is one ~24h epoch ([round] epoch_blocks): the trainer runs exactly one round per day, so the king is trained once per day and the whole day's trainings share one RoundSeeds (identical random init). Only generators committed on-chain before the epoch boundary compete in that round — commit late and you're in the next one. Each round has two stages: a cheap heat trains every eligible challenger for [round] heat_train_hours (~30min, primary size) and the owner screens them down to the top [round] finalists; the final then trains the king and the surviving finalist to the full [training] target_train_hours (~3h) at every configured size (the 4M primary plus each [[training.sizes]], e.g. 22M).

The central invariant: in a round, the king's generator and the challenger's generator are trained into models under a byte-identical contract at each size: the same Toto2 architecture and random initialisation, same compute budget, optimiser, generation seed, and training seed. The only thing that differs is the generator code. So the downstream eval is a controlled measurement of data quality, not a confound of data + luck + hyperparameters. Because the run starts from noise, the contract pins the whole recipe (see chain.toml [training], with per-size overrides in [[training.sizes]]). Each model trains for a fixed wall-clock budget (~3h on the owner's reference GPU), enforced as a fixed token count (hours × reference throughput) so king and challenger get identical compute. A raw timer would let a generator win by emitting cheap-to-step data rather than better data, and wouldn't reproduce on a re-derived audit run.

The throne is decided on the combined score across sizes: the validator pools the king-vs-finalist per-window scores from every size into one paired bootstrap, so there is a single king judged on both the 4M and 22M models (a scaling-aware KOTH, not a per-size leaderboard). A challenger takes the throne by winning dethrone_cp round(s) by a confidence-bounded margin (paired bootstrap LCB clears the win margin). The shipped chain.toml sets dethrone_cp = 1 with a flat, no-tenure margin (win_margin_start == win_margin_end, margin_warmup_rounds = 0), so a single decisive round dethrones and every king is equally challengeable; raise dethrone_cp and re-enable the warmup for the sticky, tenure-weighted variant. Weights are split equally across the current king plus up to reward_prior_kings recent distinct kings still registered (burning to burn_uid if none are), with reward_prior_kings = 0 collapsing to pure winner-take-all.

Why Toto2-4M

The fixed model is small on purpose. Toto 2.0 is the first time-series foundation family to validate a clean scaling law across its sizes (4M → 22M → 313M → 1B → 2.5B): by adopting u-μP (Maximal Update Parametrization), the learning dynamics are tuned once on the 4M model and those exact hyperparameters transfer to the 2.5B model, with predictive skill improving monotonically and without saturation as you climb the ladder (Datadog, Toto 2.0). That makes the 4M backbone the cheapest rung of a curve known to behave: it trains from scratch in hours, yet it sits on a scaling trajectory whose ordering is expected to hold as the subnet scales the fixed model up. It is also no toy: the 4M is already competitive with Toto 1.0 and Chronos-2 despite being ~30-40x smaller. A robust, predictable, inexpensive starting point is exactly what a per-round controlled experiment needs.

Technical roadmap

cascade ships the first phase of a longer program. The sequence is deliberate: prove data quality is measurable and competable before handing miners the much larger, noisier surface of training the models themselves.

• Phase 1 — Compete on data (now). The model is byte-identical; the only variable is the synthetic data generator. This is the subnet shipping today — everything else in this README describes it. The bet: better synthetic data produces better forecasters, and we can measure that cleanly per round.
• Phase 2 — Prove it scales. Show the data advantage survives model scale. µP lets hyperparameters tuned once at the 4M rung transfer up the ladder, and optimal data mixtures are roughly size-independent — so we rank the recipe cheaply at the small model and predict large-model skill before paying for it.
• Phase 3 — Open model training. Once data quality is a solved, measurable axis, widen the contract so miners compete on the models too.
• North star — multimodal. Forecasting that reads and writes across modalities: time series ↔ language ↔ vision.

Three roles

role	package	needs GPU	needs chain
miner	cascade.miner	no	to deploy
trainer (owner)	cascade.trainer	yes	to read king / sign manifest
validator	cascade.validator	yes (eval)	to set weights

Layout

cascade/
  interface/   miner-facing contract (DataGenerator ABC, output checks, static guard)
  eval/        scoring math: CRPS (MWSQL), MASE, paired bootstrap, KOTH decision
  trainer/     owner GPU service: corpus build, fixed contract, train+upload, manifest
  validator/   manifest gate, checkpoint evaluator, KOTH state machine, weights
  miner/       miner CLI: verify, deploy (push to Hippius Hub registry + commit)
  shared/      config loader, Hippius Hub registry/S3, chain client, manifest schema
  website/     the public dashboard ("notebook"): a self-contained index.html

docs/
  ARCHITECTURE.md   end-to-end flow, trust model, the controlled-experiment invariant
  INTERFACE.md      the DataGenerator submission contract for miners
  AUDIT.md          verifying published rounds with cascade-audit (receipts, tiers)
scripts/
  example_generator/   a forkable reference generator (also a test fixture)
  publish_website.py   upload the dashboard to the manifest bucket (public-read)

Console scripts

After uv sync / pip install -e .:

• cascade verify <repo_dir>: runs every check the trainer runs (layout, static guard, hash-locked deps, and the determinism check: your generator must produce a byte-identical corpus at a fixed seed).
• cascade deploy <repo_dir> --hub-repo <ns/name> --wallet-name ... --wallet-hotkey ...: verifies the local generator, pushes it to the Hippius Hub registry (OCI), and commits metro-v1:gen:hippius:<repo>@<digest> on-chain (the OCI digest pins the content — no git SHA).
• cascade-trainer --trainer cascade.trainer.toto2_trainer:Toto2Trainer: the owner training service (--offline for a config/seed smoke); the reference Toto2-4M backend lives in cascade.trainer.toto2_trainer. Add --remote-hosts hosts.toml to train the king and challenger in parallel on separate SSH GPU pods (Lium/Targon); see scripts/remote_hosts.example.toml.
• cascade-train-worker: the per-pod worker the remote dispatch runs (trains one role, uploads its checkpoint, prints a receipt — no wallet on the pod).
• cascade-validator: the validator loop (--offline for a state smoke).
• cascade-audit latest / cascade-audit round <id>: third-party verification of a published round receipt — re-derives seeds, digests, the KOTH verdict, and (at --tier 1) each generator's corpus, with a nonzero exit on any mismatch (CI-usable). No wallet or GPU needed for tiers 0–1; see docs/AUDIT.md.

Storage is Hippius: models/checkpoints/generators on the Hippius Hub registry (OCI, pinned by repo@digest), manifests + training logs on Hippius S3. Install the extra (pip install -e '.[hippius]') and set the env credentials (HIPPIUS_S3_ACCESS_KEY / HIPPIUS_S3_SECRET_KEY, and a Hub token HIPPIUS_HUB_TOKEN or HIPPIUS_HUB_USERNAME + HIPPIUS_HUB_PASSWORD).

Public round receipts. After each round's weights are set, the validator publishes a signed RoundReceipt to the manifest bucket — the full public record of the round (chain context, the trainer's manifest verbatim, the participant set, every per-window score, the KOTH verdict, and the weight vector), so a third party can re-derive the owner's work without trusting it. The layout mirrors the manifests:

s3://<manifest_bucket>/manifests/round-<id>.json   the trainer's signed manifest
s3://<manifest_bucket>/manifests/latest.json       pointer to the newest manifest
s3://<manifest_bucket>/receipts/round-<id>.json    the validator's signed receipt
s3://<manifest_bucket>/receipts/latest.json        pointer to the newest receipt
s3://<manifest_bucket>/receipts/index.json         rolling round summary (dashboard)

The dashboard ("notebook"). A single self-contained cascade/website/index.html renders the live king-of-the-hill state — the reigning king generator, the reign chain, the per-round KOTH verdicts, and geomean(CRPS·MASE) skill over time — in a paper-notebook style. It is a static page that reads only the public-read receipts above: receipts/latest.json for the current round's detail and receipts/index.json for history. The validator maintains that index (a rolling window of compact per-round summaries with a pointer back to each signed receipt) alongside every receipt it publishes — presentational only, so a stale index never affects weights, and audit trust still flows through the signed per-round receipts. Serve the page from the same bucket with python scripts/publish_website.py (needs the HIPPIUS_S3_* credentials); it then lives at <s3_endpoint>/<manifest_bucket>/index.html.

<id> is the round id (the base seed derived from the epoch-boundary block hash). A round the validator rejected still gets a receipt ("status": "rejected") carrying the gate's reason. Verify one with cascade-audit latest — see docs/AUDIT.md.

Before launching, set chain.toml [subnet] netuid, [training] base_arch_digest (sha256 of the frozen base architecture), [manifest] trainer_hotkey, [eval] window_pool (the held-out pool's Hub repo@digest), and [storage] endpoints.

Quick start

pip install -e .                 # core: numpy + scipy only
pip install -e '.[dev]'          # + pytest/ruff/hypothesis
python -m pytest tests/unit -q   # CPU tests, no torch/HF/chain needed

The heavy stacks are optional extras, pulled in only where needed: .[train] (torch/transformers for the trainer + validator evaluator), .[hippius] (Hippius Hub registry + S3 + huggingface_hub), and .[chain] (bittensor).

The Toto2-4M from-scratch training sits behind the cascade.trainer.contract.BaseTrainer protocol (the GPU boundary). A runnable reference implementation ships in cascade.trainer.toto2_trainer (a causal patch transformer with a 9-quantile pinball head, trained from random init under the chain.toml [training] recipe); it needs a GPU to validate end-to-end, so run it on your reference box before pinning base_arch_digest. Everything above that boundary is numpy/CPU and tested. See docs/ARCHITECTURE.md.

License

MIT