# Teutonic SN3

> Bittensor subnet 3. Miners train a challenger LLM that beats the
> current "king" on held-out per-token NLL. Validators decide each
> duel via a paired bootstrap test on tokenized Hippius shards.
> Winner takes 100% of SN3 emission until dethroned.

The king's architecture changes between chain versions (currently
Teutonic-XXIV). Always read live state and the current chain's mining
guide before training.

## Live state
- [dashboard.json](https://teutonic.ai/dashboard.json): king repo + revision SHA, queue, history, accepted/rejected stats. The `king` field tells you exactly what to clone.
- [Hippius dataset bucket](https://s3.hippius.com/teutonic-sn3/): tokenized shards used by validators for scoring.

## Mining
- [Mining guide (HTML)](https://teutonic.ai/mining.html): full step-by-step recipe for the current chain - environment, exact architecture lock, training script, submission flow.
- [Mining guide (markdown)](https://teutonic.ai/mining): same content as plain markdown - friendlier for LLM agents.
- [Validator + miner source](https://github.com/unarbos/teutonic): canonical algorithm. Read `teutonic/validator.py::validate_challenger_config` for the exact contract your repo must satisfy, and `teutonic/scripts/mining/train_challenger.py` for a working training harness.

## Critical constraints (validator will reject otherwise)
- Challenger MUST share `model_type` and every key in `CONFIG_MATCH_KEYS` with the king (vocab, dims, MoE shape, RoPE, etc).
- Your HF repo name MUST embed the first 8 ss58 chars of your coldkey (anti-impersonation gate, added 2026-04-29).
- Repo MUST NOT contain `*.py` files or set `config.auto_map`. Vendored modeling code only.
- Currently `attn_implementation="eager"` is the only supported path for Quasar layers (FA2 / SDPA not yet wired in upstream transformers).
- Trainability probe runs one SGD step on submission; NaN/non-finite loss = rejection.

## Get a GPU
- [Lium](https://lium.io): on-demand RTX4090 to B200, $0.12-$32/hr/GPU. Quickest to start.
- [Targon](https://targon.com): production-grade B200 boxes, ~$40/hr each.
- B200 needs the cu128 PyTorch wheel (`pip install --index-url https://download.pytorch.org/whl/cu128 torch`) - cu124 will load weights then crash with "no kernel image is available".

## Optional
- [Bittensor docs](https://docs.bittensor.com): wallet creation, hotkey registration, alpha emissions.
- [HuggingFace docs](https://huggingface.co/docs): model hosting, revision pinning.
- [SILX Quasar reference](https://huggingface.co/silx-ai/Quasar-3B-A1B-Preview): the small-scale Quasar paper architecture.