TAO—Block—Subnets85 / 128active
Affine SN120: miners upload one free batch file; the validator trains one step from the head and keeps it only if the held-out gate is GREEN (gain > 0, z > 1 flat bar, pooled cap-hit < 0.33, >= 100 tasks scored; prior-v1 baseline). GREEN files are public. Docs: affine.io/llms.txt
what the project says about itself — the input for semantic labels
Affine SN120: miners upload one free batch file; the validator trains one step from the head and keeps it only if the held-out gate is GREEN (gain > 0, z > 1 flat bar, pooled cap-hit < 0.33, >= 100 tasks scored; prior-v1 baseline). GREEN files are public. Docs: affine.io/llms.txt
scope of the data
Taonets fetches the repo's public metadata (description, topics, language, stars, last push) and the head of its README. Commit-count histories, contributor lists and release notes require the GitHub commits/releases endpoints, which the worker doesn't consume yet — those panels are coming with the ingest upgrade and are not simulated in the meantime.
first lines of the default branch README, unedited
# Affine epoch subnet Bittensor subnet 120 (MiMo GRPO). **Miners upload one free batch file** (prompts, rollouts, rewards) to a private per-hotkey object; at the miner's turn the validator copies it, checks the format, trains one GRPO step from the published head (trainer v8.3.3l) and the promotion gate keeps the candidate only if it improves the model on a private held-out set. Rewards are used as given; no mining software is needed. **Gate (prior-v1 baseline):** GREEN iff gain > 0, z > 1 (flat bar, no time relaxation), pooled cap-hit < 0.33 and at least 100 tasks scored. GREEN -> the candidate becomes the head and the turn earns; RED -> the head is kept and the turn earns nothing. **GREEN files are public** byte-exact at `https://affine.io/k1/turns/<epoch>/batch.json` (with their lr / kl in `turn.json`); RED files and RED training params stay private. - **[Mining: the batch file, upload and rules](docs/MINER_QUICKSTART.md)** · reference helper [`miner/affine_batch.py`](miner/affine_batch.py) · [miner kit and validator code](miner/README.md) · live contract https://affine.io/mining.json (`batch_file`) · https://affine.io/llms.txt - Optional per-turn `lr` (1.25e-7 to 1e-6, default 5e-7) and `kl` (0.005 to 0.1, default 0.02) in an `affine-batch-v2` file: [format and math](docs/MINER_QUICKSTART.md#optional-affine-batch-v2-optional-lr-and-kl-from-2026-10-31) · [validator side](ops/ff_train_params/README.md) - Optional per-turn `step_mult` (integer: 1, 2, default 1; duongna's step multiplier) and optimizer / loss knobs (per-group lr scales, adam_eps, weight_decay, clamp_mode, loss_agg, adv_center, adv_clip; all default to the previous trainer) in an `affine-batch-v2` file, from e575: [fields and math](docs/MINER_QUICKSTART.md) · [validator side (trainer v8.3.3l, replay, activation/revert)](ops/step_mult/README.md) - [Held-out gate and prior-v1 baseline (code)](priors/gate/README.md) - [GREEN-file publishing (default on)](ops/green_publish_default/README.md) - [Validator side of the free-file protocol (lifecycle, activation/revert)](ops/free_file/README.md) - [STANDING RULES (ops): nothing ever on the encrypted disk](STANDING-RULES.md) - [Live registration, payouts and rollback](docs/LIVE_SUBNET.md) - [Changelog](CHANGELOG.md) History and inactive designs (not the live protocol): [architecture and runtime](docs/ARCHITECTURE.md) · [MiMo operator lifecycle](docs/MIMO_LIFECYCLE.md) · [flip-incentive shadow scoring](docs/FLIP_INCENTIVE.md) · [fixed-nonce GRPO cutover plan](docs/MIMO_FIXED_NONCE_GRPO_CUTOVER_PLAN.md) · [retained-Adam NLL experiment](docs/EXPERIMENTAL_RETAINED_NLL.md) · [RCore adapter](docs/NATIVE_RCORE_COMMON.md) · [implementation plan](IMPLEMENTATION_PLAN.md) · [operational state and evidence](STATE.md) Run `.venv/bin/python -m subnet.mock` for a real model/R2/training smoke, or `.venv/bin/python -m subnet.tests` and `.venv/bin/python -m ops.new_subnet_checks` for protocol checks. Credentials and generated artifacts stay outside Git. Read AGENTS.md before modifying this repo. Legacy production and its private archive remain separate; this checkout does not automatically deploy changes. Independent, read-only evidence checks for retained experiments: ```bash .venv/bin/python -m ops.check_epoch_evidence .venv/bin/python -m ops.check_service_evidence ``` The first checks completed multi-environment research epochs. The second checks the continuous controller's deadline, signed frozen artifacts, independently accepted batches, changed weight bytes, comparable held-out evaluations and the next epoch's checkpoint binding. Both reject payable test records and confirm their scores are excluded from payout aggregation; neither submits chain writes. Authority-scoped checkpoint descriptors preserve independent controller histories. Missing after-training evaluation files remain pending rather than counting as completed epochs. These checks establish the recorded runs, not general model improvement or completion of all environment integrations. Isolated signed GPU role controls can be inspected with: ```bash .venv/bin/python -m ops.check_gpu_role_evidence --state /private/operator-gpu-control-state ``` This binds signed jobs to operator-collected reports, verifies frozen R2 bytes and checkpoint handover, and requires an independent published checkpoint byte audit. It does not itself recompute inference or prove continuous controller integration.