Take the crown.
Albedo runs on Bittensor subnet 97. If you can train a 4B coding model that beats the king, the throne is yours — permissionlessly. Here's the path.
Get the king & the data
Pull the current champion and the published winning duel traces — your raw training material.
python scripts/collect_traces.py --out data/traces.jsonlTrain your challenger
Fine-tune a Qwen3-4B on the winning trajectories. Same size class as the king is required and enforced.
python scripts/train_sft.py --base Qwen/Qwen3-4B --data data/traces.jsonl --output ckpt/Sanity-check the format
One clean solution per turn, no injected verdicts. A format fail scores zero from every judge — never skip it.
python scripts/sanity_check.py ckpt/Upload to Hippius
Push weights to decentralized storage, content-addressed by sha256 digest.
python scripts/upload_challenger.py --model ckpt/ --repo you/albedo-qwen3-4b-v1 --hotkey <ss58>Reveal on-chain
Commit your reveal signed by your hotkey. The validator picks it up in ~30s and queues your duel.
v4|you/albedo-qwen3-4b-v1|sha256:… → subtensor.commit(wallet, 97, reveal)Clear the bar. Survive the math.
Same size class
Architecture is locked to the king and checked before any GPU time — you win by training a better 4B model, not a bigger one.
No duplicates
A near-copy of an existing model (weight cosine similarity ≥ 0.95) is rejected. You can't just resubmit the king.
No gaming the judges
Every challenger passes a prompt-injection probe. Try to inject a fake verdict and you're out, fail-closed.
One eval per hotkey
Each hotkey gets one shot per model, and the reigning king can't challenge itself.
Win condition
Beat the king by ≥ 2.0 points on a 0–100 scale, and pass a 95%-confidence bootstrap test.
The margin gate stops within-noise wins. The confidence gate stops lucky ones. Both must pass to crown.