CMF-2
50.5M dual-stream model for 15-minute Polymarket crypto binaries.
Fast tape is Binance USDT-M. Slow tape is the Polymarket CLOB. The policy buys a token only when calibrated P(resolve UP) clears the displayed offer.
Code: github.com/shubhxho/polymarket-model
Card: shubhxho.github.io/polymarket-model
Windows: shubhxho/cmf-15m-binance
Files
| file | what |
|---|---|
fusion.safetensors |
CMF-2 weights, 202 MB F32 |
model.json |
architecture + train config |
metrics.json |
last held-out eval |
cmf2-hero.png / cmf2-eval.png / cmf2-arch.png |
launch cards |
Held-out simulator (80 episodes)
Same book for every policy. Share PnL after paying bid/ask. Not live Polymarket fills.
| policy | mean PnL | Sharpe | trades/ep | P(UP) |
|---|---|---|---|---|
| CMF-2 | +1.13 | 2.94 | 1.24 | 92.3% |
| lag oracle | +1.76 | 4.19 | 0.99 | — |
| random | −7.40 | −44.4 | 60.7 | — |
Train was 3.57 hours on Apple Silicon MLX (chronological split, EMA, val/best-ckpt, utility fine-tune, temperature scale). About 64% of oracle PnL.
Load
uv pip install huggingface_hub
hf download shubhxho/cmf-2 --local-dir checkpoints
from pathlib import Path
from cmf.io import load_bundle
model, cfg = load_bundle(Path("checkpoints"))
Paper is default. Live CLOB V2 needs CMF_LIVE=1 and an explicit arm.
What this is not
Planted-lag simulator. Assumes you can hit the quote you see. Not a live Sharpe and not a claim that 92% expiry accuracy transfers to production 15-minute markets.
Quantized