Incentivizing the improvement of the algorithms behind a pretraining run
Neuron registration trends
The alpha price in τ, day by day, and what moved it each way.
Log change over the last 30 days −14.63%: protocol buys 0.00%, root sells −0.86%, users net −13.24%, cross-subnet −0.52%, unattributed −0.0033%.
The chart follows the alpha price in τ. Each period starts at its opening price and steps through the mechanisms below, in a fixed order, to its closing price: a step that rises pushed the price up, one that falls pushed it down, and its height is how far it moved the price. A thin line carries each close into the next period’s open; it is dashed where the two differ. The faint candle behind each period is the market’s own open, high, low and close, for context.
With many periods on screen (90 days, or hours), the steps would be too thin to read, so the chart draws the price as a line instead and colours each period by the mechanism that moved it most. Hover a period for the full step-by-step breakdown.
Not here, because they do not move the spot price: the per-block TAO injection into the pool, and burns.
All figures are log changes, not simple percentage changes, so the parts add up exactly to the total; on big moves the two differ (a 10% log fall is a 9.5% price fall).
2026-11-04 UTC, as of ≈00:20 UTC (block 9,428,998)
The day isn’t over yet, so these figures keep changing. They are in no daily total, average or rank.
Training algorithms and data selection decisions drive model performance, yet research on these components remains siloed and difficult to validate at scale. Individual researchers lack infrastructure to test and publish improvements, and the field has no incentive mechanism to surface practical discoveries. A network of distributed researchers can accelerate this work by submitting candidate improvements to a shared evaluation harness, where validators independently test them under controlled conditions and publish results.
Subnet 125 implements this as an incentivized research platform. Miners write Python code for optimizer updates (Challenge 1) or data-selection policies (Challenge 2) and submit source + optional weights. Validators train and score both candidate and identity baseline on a fixed model (360M Llama-style), corpus (FineWeb-Edu), and compute budget (20 hours on a B200). Improvements must clear an adaptive bar (≥0.03 nats for optimizers, ≥0.02 suite points for refiners), confirmed by a second validator-funded run.
Verified improvements earn credit that halves every 14 days; frontier events determine validator weights and miner emission.
This mechanism turns research into a durable, auditable artifact: source code, training logs, checkpoint URIs, and per-item scores are published in signed round files. Any researcher can download a winning refiner or optimizer, inspect its weights and logic, reproduce results, or build on it. The subnet's global talent pool and compute resources make early gains cheaper than isolated lab work.
Auto-generated from the GitHub repository and may not reflect the latest changes.
Refinery (Bittensor subnet 125) is an incentivized research network for improving language model training algorithms. It connects miners who submit optimizer and data-refinement code to validators who evaluate submissions under fixed compute budgets. Miners are rewarded for verified improvements in validation loss (Challenge 1: optimizers) and benchmark performance (Challenge 2: refiners that decide what the model learns from each document).
Refinery uses a commit/reveal protocol where miners register a hotkey, commit a hash of their submission (source code + weights), and reveal the actual source after the commit window closes. Validators run miners' optimizer or refiner code on a fixed evaluation contract (360M model, FineWeb-Edu corpus, pinned budget, benchmark suite). Each submission is wrapped in a sandbox with restricted namespaces, process isolation, and per-token timeout.
Baseline and submission are trained side-by-side under identical conditions and scored against a curated benchmark suite. Results are published in signed round files with full audit trails. The frontier tracks confirmed improvements per challenge, applying half-life credit decay and adaptive bars.
Two pools (optimizer and refiner) blend their frontiers 50/50 for final validator weights.
Challenge 1: Submission beats frontier by >0.03 nats (floor), confirmed by second validator run at worse score. Challenge 2: Submission beats frontier by >0.02 suite points (floor, ~6.7 sd of noise), confirmed by second run. Confirmed improvements earn credit g_i × 0.5^((t−t_i)/14d).
Pool pays min(1, Σ credit / R(t)) pro rata; R ramps 0.02→0.07 over 56 days. Unearned emission burns. Each challenge holds 50% of miner emission; validator weights are 0.5×optimizer frontier + 0.5×refiner frontier.
The sections above were auto-generated from the GitHub repository and may not reflect the latest changes. View repository (github.com, opens in a new tab, external site)
Incentivizing the improvement of the algorithms behind a pretraining run
This information was auto-extracted from the GitHub repository and may be incomplete or out of date. Always refer to the subnet's GitHub repository (github.com, opens in a new tab, external site) for the latest instructions.
Not yet verified: Network
GPU not required to serve submission; meaningful local experiments require suitable compute. Miners submit algorithm source code, not compute resources.
Not yet verified: GPU model, VRAM, GPU count, Storage, Network, Notes
Machine-extracted from the subnet's repository docs; may be wrong or out of date. Check the repository before buying or renting hardware. Open the repository (github.com, opens in a new tab, external site) · Extracted 2 Nov 2026
Create and activate a dedicated Python 3.12 virtual environment
python3 -m venv .venv && . .venv/bin/activate
Install dependencies from requirements.txt
python -m pip install -r requirements.txt
Verify installation and check available commands
python -m sn125 --help
Read mining guide, implement optimizer or refiner, test locally before paid submission
Follow MINING.md to build and test a candidate submission
Serve submission through registered hotkey with commit/reveal flow
python -m sn125 mine --optimizer optimizer.py (or --refiner refiner.py --weights weights.safetensors --weights-url https://...)
Create and activate a dedicated Python 3.12 virtual environment
python3 -m venv .venv && . .venv/bin/activate
Install dependencies
python -m pip install -r requirements.txt
Build document index for FineWeb-Edu corpus (run once per box as root)
export SN125_FINEWEB_DIR=/path/to/fineweb && python -m sn125 refiner-index --data-dir $SN125_FINEWEB_DIR
Build benchmark suite (run once per box as root)
python -m sn125 bench-build --data-dir $SN125_FINEWEB_DIR --jsonl-dir data/pilot_encdec_v0/bench
Start validator with automatic updates from git; supervisor applies releases at round boundaries only
python -m sn125 validate --auto-update
| Day (UTC) | Share |
|---|---|
| 2026-11-03 | 0.0011% |
| 2026-11-02 | 0.0013% |
| 2026-11-01 | 0.0013% |
| 2026-10-31 | 0.0014% |
| 2026-10-30 | 0.0013% |
| 2026-10-29 | 0.0013% |
| 2026-10-28 | 0.0014% |
| 2026-10-27 | 0.0015% |
| 2026-10-26 | 0.0016% |
| 2026-10-25 | 0.0018% |
| 2026-10-24 | 0.0023% |
| 2026-10-23 | 0.0029% |
| 2026-10-22 | 0.0036% |
| 2026-10-21 | 0.0025% |
| 2026-10-20 | 0.0007% |
| 2026-10-19 | 0.0023% |
| 2026-10-18 | 0.0025% |
| 2026-10-17 | 0.0034% |
| 2026-10-16 | 0.0044% |
| 2026-10-15 | 0.01% |
| 2026-10-14 | 0.01% |
| 2026-10-13 | 0.01% |
| 2026-10-12 | 0.01% |
| 2026-10-11 | 0.01% |
| 2026-10-10 | 0.01% |
| 2026-10-09 | 0.01% |
| 2026-10-08 | 0.01% |
| 2026-10-07 | 0.01% |
| 2026-10-06 | 0.02% |
| 2026-10-05 | 0.01% |
| 2026-10-04 | 0.01% |
| 2026-10-03 | 0.01% |
| 2026-10-02 | 0.02% |
| 2026-10-01 | 0.01% |
| 2026-09-30 | 0.02% |
| 2026-09-29 | 0.01% |
| 2026-09-28 | 0.01% |
| 2026-09-27 | 0.01% |
| 2026-09-26 | 0.02% |
| 2026-09-25 | 0.02% |
| 2026-09-24 | 0.02% |
| 2026-09-23 | 0.02% |
| 2026-09-22 | 0.02% |
| 2026-09-21 | 0.02% |
| 2026-09-20 | 0.02% |
| 2026-09-19 | 0.02% |
| 2026-09-18 | 0.02% |
| 2026-09-17 | 0.02% |
| 2026-09-16 | 0.02% |
| 2026-09-15 | 0.02% |
| 2026-09-14 | 0.02% |
| 2026-09-13 | 0.0047% |
| 2026-09-12 | 0.00% |
| 2026-09-11 | 0.00% |
| 2026-09-10 | 0.00% |
| 2026-09-09 | 0.00% |
| 2026-09-08 | 0.00% |
| 2026-09-07 | 0.00% |
| 2026-09-06 | 0.00% |
| 2026-09-05 | 0.00% |
| 2026-09-04 | 0.00% |
| 2026-09-03 | 0.00% |
| 2026-09-02 | 0.00% |
| 2026-09-01 | 0.00% |
| 2026-08-31 | 0.00% |
| 2026-08-30 | 0.00% |
| 2026-08-29 | 0.00% |
| 2026-08-28 | 0.00% |
| 2026-08-27 | 0.00% |
| 2026-08-26 | 0.00% |
| 2026-08-25 | 0.00% |
| 2026-08-24 | 0.00% |
| 2026-08-23 | 0.00% |
| 2026-08-22 | 0.00% |
| 2026-08-21 | 0.00% |
| 2026-08-20 | 0.00% |
| 2026-08-19 | 0.00% |
| 2026-08-18 | 0.00% |
| 2026-08-17 | 0.00% |
| 2026-08-16 | 0.00% |
| 2026-08-15 | 0.00% |
| 2026-08-14 | 0.00% |
| 2026-08-13 | 0.00% |
| 2026-08-12 | 0.00% |
| 2026-08-11 | 0.00% |
| 2026-08-10 | 0.00% |
| 2026-08-09 | 0.00% |
| 2026-08-08 | 0.00% |
| 2026-08-07 | 0.00% |
| 2026-08-06 | 0.00% |
