Tenzro Testnet is live —request testnet TNZO
← Back to tutorials

Post a training task on Tenzro Train

This tutorial walks the sponsor flow end to end: write a TimesFM-class training task spec, escrow TNZO into the reward pool, watch trainers enroll, monitor round progress, and fetch the sealed TrainingReceipt when the run completes. Phase 1 ships timeseries-first on the Open tier with Mean aggregation — those are the defaults this tutorial uses.

Prerequisites

1. Write the task spec

A TrainingTaskSpec is a single JSON document that fully describes the run. The spec is committed verbatim into the final receipt, so every field — sponsor DID, dataset hash, hyperparameters — becomes an audit-trail anchor. Save this as timesfm-task.json:

{
  "task_id": "train-2026-04-25-timesfm-200m",
  "sponsor_did": "did:tenzro:human:01J2Z7K8...",
  "sponsor_address": "0x7a4bcb13a6b2b384c284b5caa6e5ef3126527f93",
  "architecture": {
    "family": "timesfm",
    "param_count": 200000000,
    "modality": "Timeseries",
    "fragment_count": 12,
    "dtype": "bf16",
    "metadata": {
      "context_len": 2048,
      "horizon": 128,
      "n_quantiles": 10
    }
  },
  "tier": "Open",
  "aggregation": "Mean",
  "trainer_count": 8,
  "quorum": 6,
  "inner_steps": 24,
  "max_rounds": 1024,
  "grace_window_ms": 30000,
  "reward_pool": "1000000000000000000000",
  "dataset_ref": "ipfs://QmXyZ...",
  "dataset_hash": "0xc0ffee...deadbeef",
  "min_throughput": null,
  "created_at": 1761350400,
  "metadata": {
    "eval_suite": "monash-tsf-v2",
    "target_mase": 0.8
  }
}

Key fields to understand:

2. Post the task

The CLI reads the spec, signs it with your wallet, and dispatches to tenzro_training_postTask. The reward pool is escrowed atomically — the call fails if your wallet is underfunded.

tenzro train post-task --spec ./timesfm-task.json \
  --rpc https://rpc.tenzro.network

The same call over raw JSON-RPC, if you're integrating from another tool:

curl https://rpc.tenzro.network \
  -X POST \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tenzro_training_postTask",
    "params": {
      "task_spec": <CONTENTS OF timesfm-task.json>
    }
  }' | jq

# Expected:
# {
#   "jsonrpc": "2.0",
#   "id": 1,
#   "result": {
#     "task_id": "train-2026-04-25-timesfm-200m",
#     "status": "awaiting_enrollment"
#   }
# }

On success, the run enters awaiting_enrollment. The syncer is elected (VRF-weighted by stake), and the gossip topics tenzro/training/1.0.0 and tenzro/training/syncer/1.0.0 begin advertising the new run.

3. Watch enrollment

Trainers stake into the run until trainer_count slots fill. Once the run reaches the quorum threshold and the syncer accepts the roster, status flips to running:

# Poll the run state — wait for status to flip from
# awaiting_enrollment → running once K=6 trainers stake in.
while true; do
  STATUS=$(tenzro train get-run \
    --task-id train-2026-04-25-timesfm-200m \
    --rpc https://rpc.tenzro.network \
    --format json | jq -r '.status')
  ENROLLED=$(tenzro train get-run \
    --task-id train-2026-04-25-timesfm-200m \
    --rpc https://rpc.tenzro.network \
    --format json | jq -r '.enrolled_trainers')

  echo "$(date +%T) status=$STATUS enrolled=$ENROLLED/8"
  [ "$STATUS" = "running" ] && break
  sleep 10
done

4. Monitor round progress

Once running, the syncer finalizes one round at a time. Each round produces a fresh run_root — the SHA-256 Merkle root over every accepted outer gradient, prefixed with tenzro/train/run-root/v1.

# Once running, the syncer commits one round at a time.
# Watch current_round march toward max_rounds=1024.
curl -s https://rpc.tenzro.network \
  -X POST \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 2,
    "method": "tenzro_training_getRun",
    "params": { "task_id": "train-2026-04-25-timesfm-200m" }
  }' | jq '.result | {status, current_round, max_rounds, run_root}'

For real-time visibility, subscribe to the tenzro/training/syncer/1.0.0 gossip topic instead of polling.

5. Fetch the receipt

When current_round reaches max_rounds, the syncer seals a TrainingReceipt and stores it in CF_TRAINING_RECEIPTS. Any unspent escrow is refunded to sponsor_address.

tenzro train get-receipt \
  --task-id train-2026-04-25-timesfm-200m \
  --rpc https://rpc.tenzro.network

The receipt shape:

{
  "task_id": "train-2026-04-25-timesfm-200m",
  "task_spec": { /* full original spec, bound verbatim */ },
  "syncer_did": "did:tenzro:machine:syn1...",
  "trainers": ["did:tenzro:machine:t1...", ...],
  "rounds_completed": 1024,
  "final_model_hash": "0xa1b2c3...",
  "run_root": "0xd4e5f6...",
  "aggregation_transcript": "0x...",
  "sealed_at": 1761436800,
  "syncer_signature": "0x..."
}

The receipt is verifiable on-chain via the TRAINING_VERIFY precompile at 0x1008. Any contract — including ERC-20 reward escrows — can gate payouts on a receipt without trusting the sponsor.

Next steps